Recent advances in deep learning for object detection

May combine two claims.

From Table 3, most state-of-the-art detection algorithms are based on classification backbones, and only a few of them try different selections such as CornerNet based on Hourglass Net.

“From Table 3, most state-of-the-art detection algorithms are based on classification backbones, and only a few of them try different selections (such as CornerNet based on Hourglass Net).”

May combine two claims.

Traditional detectors generated a huge number of proposals during proposal generation, many redundant, resulting in a large number of false positives during classification.

“during proposal generation, a huge number of proposals were generated, and many of them were redundant; this resulted in a large number of false positives during classification.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Instance segmentation is more sensitive to spatial misalignment, and thus has higher requirement to process the spatial information.

“Due to more precise pixel-level prediction, instance segmentation is more sensitive to spatial misalignment, and thus has higher requirement to process the spatial information.”

May combine two claims.

In traditional detectors, feature descriptors were hand-crafted based on low level visual cues, which made it difficult to capture representative semantic information in complex contexts.

“feature descriptors were hand-crafted based on low level visual cues [31, 32, 23], which made it difficult to capture representative semantic information in complex contexts.”

• cites [31] [31] D. G. Lowe, Distinctive image features from scale-invariant keypoints, in: IJCV, 2004. -Not checked yet

  Not checked in the free run.

• cites [32] [32] T. Ojala, M. Pietikainen, T. Maenpaa, Multiresolution gray-scale and rotation invariant texture classification with local binary patterns, in: TPAMI, 2002. -Not checked yet

  Not checked in the free run.

• cites [23] [23] H. Bay, T. Tuytelaars, L. Van Gool, Surf: Speeded up robust features, in: ECCV, 2006. -Not checked yet

  Not checked in the free run.

May combine two claims.

In traditional detectors, window scales were designed manually and heuristically, and could not match the objects well.

“Moreover, window scales were designed manually and heuristically, and could not match the objects well”

May combine two claims.

A deep convolutional neural network can obtain a better feature representation with a larger dataset, while the learning capacity of traditional visual descriptors are fixed, and can not improve when more data becomes available.

“a deep convolutional neural network can obtain a better feature representation with a larger dataset, while the learning capacity of traditional visual descriptors are fixed, and can not improve when more data becomes available.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

R-CNN is a pioneering two-stage object detector proposed by Girshick et al. in 2014.

“R-CNN [2] is a pioneering two-stage object detector proposed by Girshick et al. in 2014.”

• cites [2] [2] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for accurate object detection and semantic segmentation, in: CVPR, 2014. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Due to the tremendous successes of deep learning based image classification, object detection techniques using deep learning have been actively studied in recent years.

“Due to the tremendous successes of deep learning based image classification, object detection techniques using deep learning have been actively studied in recent years.”

May combine two claims.

Compared to the previous state-of-the-art methods based on a traditional detection framework SegDPM with 40.4% mAP on Pascal VOC2010, RCNN significantly improved the detection performance and obtained 53.7% mAP.

“Compared to the previous state-of-the-art methods based on a traditional detection framework SegDPM [44] with 40.4% mAP on Pascal VOC2010, RCNNsignificantly improved the detection performance and obtained 53.7% mAP.”

• cites [44] [44] S. Fidler, R. Mottaghi, A. Yuille, R. Urtasun, Bottom-up segmentation for top-down detection, in: CVPR, 2013. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

In traditional detectors, each step of the detection pipeline was designed and optimized separately, and thus could not obtain a global optimal solution for the whole system.

“each step of the detection pipeline was designed and optimized separately, and thus could not obtain a global optimal solution for the whole system.”

May combine two claims.

Deep convolutional neural networks generate hierarchical feature representations from raw pixels to high level semantic information, which is learned automatically from the training data and shows more discriminative expression capability in complex contexts.

“deep convolutional neural networks generate hierarchical feature representations from raw pixels to high level semantic information, which is learned automatically from the training data and shows more discriminative expression capability in complex contexts.”

May combine two claims.

Two-stage detectors commonly achieve better detection performance and report state-of-the-art results on public benchmarks.

“Two-stage detectors commonly achieve better detection performance and report state-of-the-art results on public benchmarks”

• Also stated in: § 3.2. Detection Paradigms · p. 6 · plain

  “Two-stage detectors often reported state-of-the-art results on many public benchmark datasets.”

Probably not a substantial claim: may not pass S4 (based).

The new deep learning based algorithms outperformed the traditional detection algorithms by huge margins.

“The new deep learning based algorithms outperformed the traditional detection algorithms by huge margins.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Though the two detection settings are slightly different, the main components introduced later can mostly be shared by the two settings.

“Though the two detection settings are slightly different, the main components introduced later can mostly be shared by the two settings.”

May combine two claims.

After the success of applying DCNN for image classification, object detection achieved remarkable progress based on deep learning techniques.

“After the success of applying deep convolutional neural networks(DCNN) for image classification [33, 1], object detection also achieved remarkable progress based on deep learning techniques [34, 2].”

• cites [33] [33] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, in: NeurIPS, 2012. -Not checked yet

  Not checked in the free run.

• cites [1] [1] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [34] [34] S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object detection with region proposal networks, in: NeurIPS, 2015. -Not checked yet

  Not checked in the free run.

• cites [2] [2] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for accurate object detection and semantic segmentation, in: CVPR, 2014. -Not checked yet

  Not checked in the free run.

May combine two claims.

These properties made it possible to design object detection algorithms based on deep convolutional neural networks which could be optimized in an end-to-end manner, with more powerful feature representation capability.

“These properties made it possible to design object detection algorithms based on deep convolutional neural networks which could be optimized in an end-to-end manner, with more powerful feature representation capability.”

May combine two claims.

One-stage detectors are significantly more time-efficient and have greater applicability to real-time object detection than two-stage detectors.

“while one-stage detectors are significantly more time-efficient and have greater applicability to real-time object detection.”

• Also stated in: § 3.2. Detection Paradigms · p. 6 · plain

  “One-stage detectors are much faster and more desired for real-time object detection applications, but have a relatively poor performance compared to the two-stage detectors.”

May combine two claims.

RPN could be inserted into Fast R-CNN and thus the whole framework could be optimized in an end-to-end manner on training data.

“RPN could be inserted into Fast R-CNN and thus the whole framework could be optimized in an end-to-end manner on training data.”

May combine two claims.

The training of SPP-net was still multi-stage and thus it could not be optimized end-to-end (and required extra cache memory to store extracted features).

“However, the training of SPP-net was still multi-stage and thus it could not be optimized end-to-end (and required extra cache memory to store extracted features).”

Probably not a substantial claim: may not pass S3 (load-bearing).

Selective Search is designed to reject regions that can easily be identified as background regions.

“Selective Search [45], which is designed to reject regions that can easily be identified as background regions.”

• cites [45] [45] J. R. Uijlings, K. E. Van De Sande, T. Gevers, A. W. Smeulders, Selective search for object recognition, in: IJCV , 2013. -Not checked yet

  Not checked in the free run.

May combine two claims.

The proposal generation step still relied on traditional methods such as Selective Search or Edge Boxes, which were based on low-level visual cues and could not be learned in a data-driven manner.

“Despite the progress in learning detectors, the proposal generation step still relied on traditional methods such as Selective Search [45] or Edge Boxes [48], which were based on lowlevel visual cues and could not be learned in a data-driven manner.”

• cites [45] [45] J. R. Uijlings, K. E. Van De Sande, T. Gevers, A. W. Smeulders, Selective search for object recognition, in: IJCV , 2013. -Not checked yet

  Not checked in the free run.

• cites [48] [48] C. L. Zitnick, P. Doll´ ar, Edge boxes: Locating object proposals from edges, in: ECCV, 2014. -Not checked yet

  Not checked in the free run.

May combine two claims.

YOLO used only the last feature map for prediction, which was not suitable for predicting objects at multiple scales and aspect ratios.

“only the last feature map was used for prediction, which was not suitable for predicting objects at multiple scales and aspect ratios.”

May combine two claims.

R-CNN's features of each proposal were extracted separately, leading to heavily duplicated computations, making R-CNN extremely time-consuming for training and testing.

“the features of each proposal were extracted by deep convolutional networks separately (i.e., computation was not shared), which led to heavily duplicated computations. Thus, R-CNN was extremely time-consuming for training and testing”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

R-CNN is unable to enjoy the benefits of GPU acceleration.

“Moreover, it is unable to enjoy the benefits of GPU acceleration.”

May combine two claims.

SSD achieved comparable detection accuracy with Faster R-CNN but enjoyed the ability to do real-time inference.

“SSD achieved comparable detection accuracy with Faster R-CNN but enjoyed the ability to do real-time inference.”

• cites [42] [42] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, A. C. Berg, SSD: Single shot multibox detector, in: ECCV, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

He et al. proposed Mask R-CNN, which predicted bounding boxes and segmentation masks in parallel based on proposals and reported state-of-the-art results.

“He et al. [3] proposed Mask R-CNN , which predicted bounding boxes and segmentation masks in parallel based on the proposals and reported state-of-the-art results.”

• cites [3] [3] K. He, G. Gkioxari, P. Doll´ ar, R. Girshick, Mask r-cnn, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

SPP layer did not back-propagate gradients to convolutional kernels and thus all the parameters before the SPP layer were frozen.

“In addition, SPP layer did not back-propagate gradients to convolutional kernels and thus all the parameters before the SPP layer were frozen.”

May combine two claims.

OverFeat showed significant speed strength compared with RCNN by sharing computation of overlapping regions using convolutional layers, requiring only a single forward pass.

“OverFeat showed significant speed strength compared with RCNN by sharing the computation of overlapping regions using convolutional layers, and only a single pass forward through the network was required.”

Probably not a substantial claim: may not pass S3 (load-bearing).

One-stage detectors do not have a separate stage for proposal generation.

“one-stage detectors do not have a separate stage for proposal generation (or learning a proposal generation).”

May combine two claims.

RPN enabled proposal generation in a data driven manner, and was also able to enjoy the discriminative power of deep backbone networks.

“This way RPN enabled proposal generation in a data driven manner, and was also able to enjoy the discriminative power of deep backbone networks.”

May combine two claims.

FPN achieved significant progress in detecting multi-scale objects and has been widely used in many other domains such as video detection and human pose recognition.

“FPN achieved significant progress in detecting multi-scale objects and has been widely used in many other domains such as video detection [53, 54] and human pose recognition [55, 56].”

• cites [53] [53] K. Kang, W. Ouyang, H. Li, X. Wang, Object detection from video tubelets with convolutional neural networks, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [54] [54] W. Han, P. Khorrami, T. L. Paine, P. Ramachandran, M. Babaeizadeh, H. Shi, J. Li, S. Yan, T. S. Huang, Seq-nms for video object detection, in: arXiv preprint arXiv:1602.08465, 2016. -Not checked yet

  Not checked in the free run.

• cites [55] [55] M. Rayat Imtiaz Hossain, J. Little, Exploiting temporal information for 3d human pose estimation, in: ECCV, 2018. -Not checked yet

  Not checked in the free run.

• cites [56] [56] G. Pavlakos, X. Zhou, K. G. Derpanis, K. Daniilidis, Coarse-to-fine volumetric prediction for single-image 3d human pose, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

YOLO could detect up to only two objects at a given location, making it difficult to detect small and crowded objects.

“it could detect upto only two objects at a given location, which made it difficult to detect small objects and crowded objects [40].”

• cites [40] [40] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

Simply removing the fully connected layers would result in the drastic decline of detection performance, as the deep network would have reduced the spatial information of proposals.

“Simply removing the fully connected layers would result in the drastic decline of detection performance, as the deep network would have reduced the spatial information of proposals.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

In Fast RCNN, the feature extraction, region classification and bounding box regression steps can all be optimized end-to-end, without extra cache space to store features.

“In Fast RCNN, the feature extraction, region classification and bounding box regression steps can all be optimized end-to-end, without extra cache space to store features (unlike SPP Net).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Faster R-CNN was able to make predictions at 5FPS on GPU and achieved state-of-the-art results on many public benchmark datasets, such as Pascal VOC 2007, 2012 and MSCOCO.

“Faster R-CNN was able to make predictions at 5FPS on GPU and achieved state-of-the-art results on many public benchmark datasets, such as Pascal VOC 2007, 2012 and MSCOCO.”

May combine two claims.

R-CNN's three steps (proposal generation, feature extraction and region classification) were independent components and the whole detection framework could not be optimized in an end-to-end manner, making it difficult to obtain global optimal solution.

“the three steps of R-CNN (proposal generation, feature extraction and region classification) were independent components and the whole detection framework could not be optimized in an endto-end manner, making it difficult to obtain global optimal solution”

May combine two claims.

This made it difficult to detect objects at different scales. In particular, it was difficult to detect small objects.

“This made it difficult to detect objects at different scales. In particular, it was difficult to detect small objects.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

SPP-net achieved better results and had a significantly faster inference speed compared to R-CNN.

“SPP-net achieved better results and had a significantly faster inference speed compared to R-CNN.”

May combine two claims.

YOLOv2 significantly improved detection performance while maintaining real-time inference speed.

“Redmon et al. proposed an improved YOLO version, YOLOv2 [41] which significantly improved detection performance but still maintained real-time inference speed.”

• cites [41] [41] J. Redmon, A. Farhadi, Yolo9000: better, faster, stronger, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

The extracted feature vectors maintained spatial information and thus the detector achieved competitive results compared to Faster RCNN without region-wise fully connected layer operations.

“The extracted feature vectors maintained spatial information and thus the detector achieved competitive results compared to Faster RCNN without region-wise fully connected layer operations.”

May combine two claims.

YOLO considered object detection as a regression problem and spatially divided the whole image into a fixed number of grid cells.

“YOLO considered object detection as a regression problem and spatially divided the whole image into fixed number of grid cells (e.g. using a 7 × 7 grid).”

• cites [40] [40] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

One-stage detectors typically consider all positions on the image as potential objects and try to classify each region of interest as either background or a target object.

“They typically consider all positions on the image as potential objects, and try to classify each region of interest as either background or a target object.”

May combine two claims.

SPP-net was proposed to accelerate R-CNN as well as learn more discriminative features.

“He et al. proposed SPP-net [47] to accelerate R-CNN as well as learn more discriminative features.”

• cites [47] [47] K. He, X. Zhang, S. Ren, J. Sun, Spatial pyramid pooling in deep con- volutional networks for visual recognition, in: ECCV, 2014. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

This transfer of knowledge from the Imagenet dataset offers significant performance gains.

“This transfer of knowledge from the Imagenet dataset offers significant performance gains.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

R-CNN rejects huge number of easy negatives before training, which helps improve learning speed and reduce false positives.

“In addition, R-CNN rejects huge number of easy negatives before training, which helps improve learning speed and reduce false positives.”

May combine two claims.

Fast R-CNN achieved a much better detection accuracy than R-CNN and SPP-net, and had a better training and inference speed.

“Fast R-CNN achieved a much better detection accuracy than R-CNN and SPP-net, and had a better training and inference speed.”

Probably not a substantial claim: may not pass S3 (load-bearing).

OverFeat extended the original classifier into a detector by viewing the last FC layers as 1x1 convolutional layers to allow arbitrary input.

“OverFeat extended the original classifier into detector by viewing the last FC layers as 1x1 convolutional layers to allow arbitrary input.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Selective Search relied on low-level visual cues and thus struggled to generate high quality proposals in complex contexts.

“Selective Search relied on low-level visual cues and thus struggled to generate high quality proposals in complex contexts.”

May combine two claims.

YOLO could make predictions at 45 FPS and reach 155 FPS with a more simplified backbone.

“Based on a carefully designed lightweight architecture, YOLO could make prediction at 45 FPS, and reach 155 FPS with a more simplified backbone.”

• cites [40] [40] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified, real-time object detection, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Compared to traditional hand-crafted feature descriptors, deep neural networks generate hierarchical features and capture different scale information in different layers, and finally produce robust and discriminative features for classification.

“Compared to traditional hand-crafted feature descriptors, deep neural networks generate hierarchical features and capture different scale information in different layers, and finally produce robust and discriminative features for classification.”

May combine two claims.

SPP-net does not suffer from information loss and unwanted geometric distortion.

“Thus, it does not suffer from information loss and unwanted geometric distortion.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Dai et al. proposed MNC, a multi-stage algorithm that divided the detection framework into multiple stages and predicted segmentation masks from learned bounding box proposals, which were later categorized by region classifiers.

“Later, Dai et al. [59] proposed a multi-stage algorithm named 'MNC' which divided the whole detection framework into multiple stages and predicted segmentation masks from the learned bounding box proposals, which were later categorized by region classifiers.”

• cites [59] [59] J. Dai, K. He, J. Sun, Instance-aware semantic segmentation via multitask network cascades, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

In DCNN feature representations, deep layer features are semantically-strong but spatially-weak, while shallow layer features are semantically-weak but spatially-strong.

“In DCNN feature representations, deep layer features are semantically-strong but spatially-weak, while shallow layer features are semanticallyweak but spatially-strong.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

There are four main paradigms addressing multi-scale feature learning: Image Pyramid, Prediction Pyramid, Integrated Features and Feature Pyramid.

“There are four main paradigms addressing multi-scale feature learning problem: Image Pyramid, Prediction Pyramid, Integrated Features and Feature Pyramid.”

May combine two claims.

Fusing spatially rich shallow-layer features with semantic-rich deep-layer features yields constructed features that contain rich information and can detect objects at different scales.

“By fusing spatially rich shallow layer features and semantic-rich deep layer features, the new constructed features contain rich information and thus can detect objects at different scales.”

• cites [102] [102] A. Shrivastava, R. Sukthankar, J. Malik, A. Gupta, Beyond skip connections: Top-down modulation for object detection, in: arXiv preprint arXiv:1612.06851, 2016. -Not checked yet

  Not checked in the free run.

• cites [51] [51] S. Bell, C. Lawrence Zitnick, K. Bala, R. Girshick, Inside-outside net: Detecting objects in context with skip pooling and recurrent neural networks, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [50] [50] T. Kong, A. Yao, Y. Chen, F. Sun, Hypernet: Towards accurate region proposal generation and joint object detection, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [103] [103] H. Wang, Q. Wang, M. Gao, P. Li, W. Zuo, Multi-scale location-aware kernel representation for object detection, in: CVPR, 2018. -Not checked yet

  Not checked in the free run.

• cites [104] [104] K.-H. Kim, S. Hong, B. Roh, Y. Cheon, M. Park, Pvanet: deep but lightweight neural networks for real-time object detection, in: arXiv preprint arXiv:1608.08021, 2016. -Not checked yet

  Not checked in the free run.

• cites [105] [105] O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in: International Conference on Medical image computing and computer-assisted intervention, 2015. -Not checked yet

  Not checked in the free run.

May combine two claims.

Increasing model depth to 20 layers by simply stacking convolutional layers led to optimization challenges with SGD, and performance declined significantly and was inferior to shallower models even during training.

“However, increasing model depth to 20 layers by simply stacking convolutional layers led to optimization challenges with SGD. The performance declined significantly and was inferior to shallower models, even during the training stages.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Selective Search was widely used in many detection frameworks due to its efficiency and high recall compared to other traditional methods.

“Selective Search is widely used in many detection frameworks due to its efficiency and high recall compared to other traditional methods.”

• cites [45] [45] J. R. Uijlings, K. E. Van De Sande, T. Gevers, A. W. Smeulders, Selective search for object recognition, in: IJCV , 2013. -Not checked yet

  Not checked in the free run.

May combine two claims.

CornerNet modeled information of top-left and bottom-right corners with novel feature embedding methods and corner pooling layer to correctly match keypoints belonging to the same objects, obtaining state-of-the-art results on public benchmarks.

“CornerNet modeled information of top-left and bottom-right corners with novel feature embedding methods and corner pooling layer to correctly match keypoints belonging to the same objects, obtaining state-of-the-art results on public benchmarks.”

• cites [63] [63] H. Law, J. Deng, Cornernet: Detecting objects as paired keypoints, in: ECCV, 2018. -Not checked yet

  Not checked in the free run.

May combine two claims.

DenseNet retained shallow layer features and improved information flow by concatenating the input with the residual output instead of element-wise addition.

“They proposed DenseNet [71], which retained the shallow layer features, and improved information flow, by concatenating the input with the residual output instead of”

• cites [71] [71] G. Huang, Z. Liu, L. Van Der Maaten, K. Q. Weinberger, Densely connected convolutional networks., in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

Hourglass module could capture both local and global information and thus was very suitable for object detection.

“Hourglass module could capture both local and global information and thus was very suitable for object detection.”

• cites [79] [79] A. Newell, K. Yang, J. Deng, Stacked hourglass networks for human pose estimation, in: ECCV, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

MobileNet significantly reduced computation cost and number of parameters without significant loss in classification accuracy.

“MobileNet significantly reduced computation cost as well as number of parameters without significant loss in classification accuracy.”

• cites [74] [74] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H. Adam, Mobilenets: Efficient convolutional neural networks for mobile vision applications, in: arXiv preprint arXiv:1704.04861, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

ResNeXt considerably reduced computation and memory cost while maintaining comparable classification accuracy.

“Based on ResNet, Xie et al. [73] proposed ResNeXt which considerably reduced computation and memory cost while maintaining comparable classification accuracy.”

• cites [73] [73] S. Xie, R. Girshick, P. Doll´ ar, Z. Tu, K. He, Aggregated residual transformations for deep neural networks, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

ION cropped region features from different layers via ROI Pooling and combined multi-scale region features for final prediction.

“Bell et al. proposed Inside-Outside Network (ION) [51] which cropped region features from different layers via ROI Pooling [38], and combined these multi-scale region features for the final prediction.”

• cites [51] [51] S. Bell, C. Lawrence Zitnick, K. Bala, R. Girshick, Inside-outside net: Detecting objects in context with skip pooling and recurrent neural networks, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [38] [38] R. Girshick, Fast r-cnn, in: ICCV, 2015. -Not checked yet

  Not checked in the free run.

May combine two claims.

Traditional computer vision methods cannot be jointly optimized with the whole detection pipeline and are unable to exploit the power of large scale datasets to improve representation learning.

“They cannot be jointly optimized with the whole detection pipeline. Thus they are unable to exploit the power of large scale datasets to improve representation learning.”

May combine two claims.

YOLOv2 defined better anchor priors by k-means clustering from the training data instead of setting them manually, which helped reduce optimization difficulties in localization.

“inspired by the anchor strategy used in SSD, YOLOv2 defined better anchor priors by k-means clustering from the training data (instead of setting manually). This helped in reducing optimizing difficulties in localization.”

• cites [41] [41] J. Redmon, A. Farhadi, Yolo9000: better, faster, stronger, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Producing high quality segmentation masks is very time-consuming and not applicable to large scale datasets.

“However, producing high quality segmentation masks is very time-consuming and it's not applicable to large scale datasets.”

May combine two claims.

Shortcut connections create a highway that directly propagates gradients from deep layers to shallow units and thus significantly reduces training difficulty.

“Shortcut connection creates a highway which directly propagates the gradients from deep layers to shallow units and thus, significantly reduces training difficulty.”

May combine two claims.

CenterNet combined center-based and corner-based methods, first predicting bounding boxes by pairs of corners and then predicting center probabilities to reject easy negatives, and obtained significant improvements compared with baselines.

“Duan et al. [65] proposed CenterNet, which combined the idea of center-based methods and corner-based methods. CenterNet first predicted bounding boxes by pairs of corners, and then predicted center probabilities of the initial prediction to reject easy negatives. CenterNet obtained significant improvements compared with baselines.”

• cites [65] [65] K. Duan, S. Bai, L. Xie, H. Qi, Q. Huang, Q. Tian, Centernet: Keypoint triplets for object detection, in: arXiv preprint arXiv:1904.08189, 2019. -Not checked yet

  Not checked in the free run.

May combine two claims.

Classification makes predictions on a single feature map, while detection requires feature maps with multiple representations to detect objects at multiple scales.

“classification makes predictions on a single feature map, while detection requires feature maps with multiple representations to detect objects at multiple scales.”

May combine two claims.

MLKP captured high-order statistics of proposal features and generated more discriminative feature representations efficiently.

“Wang et al. proposed a novel framework Multi-scale Location-aware Kernel Representation (MLKP) [103] which captured high-order statistics of proposal features and generated more discriminative feature representations efficiently.”

• cites [103] [103] H. Wang, Q. Wang, M. Gao, P. Li, W. Zuo, Multi-scale location-aware kernel representation for object detection, in: CVPR, 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Feature normalization is required because feature norms of different layers have a high variance.

“Feature normalization is required as feature norms of different layers have a high variance.”

May combine two claims.

Anchor priors are manually designed with multiple scales and aspect ratios in a heuristic manner, and these design choices may not be optimal.

“Despite promising performance, the anchor priors are manually designed with multiple scales and aspect ratios in a heuristic manner. These design choices may not be optimal, and different datasets would require different anchor design strategies.”

May combine two claims.

VGG16 showed that increasing depth of networks by stacking convolutional layers could increase the model's expression capability and lead to better performance.

“VGG16 showed that increasing depth of networks by stacking convolutional layers could increase the model's expression capability, and led to a better performance.”

• cites [69] [69] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, in: arXiv preprint arXiv:1409.1556, 2014. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

RPN generates proposals in a supervised way based on deep convolutional feature maps.

“Ren et al. proposed Region Proposal Network (RPN) [34] to generate proposals in a supervised way based on deep convolutional feature maps.”

• cites [34] [34] S. Ren, K. He, R. Girshick, J. Sun, Faster r-cnn: Towards real-time object detection with region proposal networks, in: NeurIPS, 2015. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Singh et al. argued that learning a single scale-robust detector to handle all scale objects was much more difficult than learning scale-dependent detectors with image pyramids.

“They argued that learning a single scale-robust detector to handle all scale objects was much more difficult than learning scale-dependent detectors with image pyramids.”

• cites [98] [98] B. Singh, L. S. Davis, An analysis of scale invariance in object detection-snip, in: CVPR, 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

CornerNet detected objects as a pair of corners.

“Law and Deng proposed a novel anchor-free framework CornerNet [63] which detected objects as a pair of corners.”

• cites [63] [63] H. Law, J. Deng, Cornernet: Detecting objects as paired keypoints, in: ECCV, 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Image Pyramid resizes input images into a number of different scales and trains multiple detectors, each responsible for a certain range of scales.

“An intuitive idea is to resize input images into a number of different scales (Image Pyramid) and to train multiple detectors, each of which is responsible for a certain range of scales [98, 99, 100, 101].”

• cites [98] [98] B. Singh, L. S. Davis, An analysis of scale invariance in object detection-snip, in: CVPR, 2018. -Not checked yet

  Not checked in the free run.

• cites [99] [99] P. Hu, D. Ramanan, Finding tiny faces, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

• cites [100] [100] F. Yang, W. Choi, Y. Lin, Exploit all the layers: Fast and accurate cnn object detector with scale dependent pooling and cascaded rejection classifiers, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [101] [101] Y. Liu, H. Li, J. Yan, F. Wei, X. Wang, X. Tang, Recurrent scale approximation for object detection in cnn, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

Shallow layer features with spatial-rich information have higher resolution and smaller receptive fields and are more suitable for detecting small objects, while semantic-rich features in deep layers are more robust to illumination and translation, have larger receptive fields but coarse resolutions, and are more suitable for detecting large objects.

“shallow layer features with spatial-rich information have higher resolution and smaller receptive fields and thus are more suitable for detecting small objects, while semantic-rich features in deep layers are more robust to illumination, translation and have larger receptive fields (but coarse resolutions), and are more suitable for detecting large objects.”

May combine two claims.

Object detection requires bounding box level annotations and thus annotating a large scale detection dataset requires much more effort and time than image classification.

“Different from image classification, object detection requires bounding box level annotations and thus, annotating a large scale detection dataset requires much more effort and time(ImageNet has 1000 categories for image classification while only 200 of them have detection annotations).”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Precise localization is challenging because predictions are commonly focused on the most discriminative part of the objects, and not necessarily the region containing the object.

“Precise localization is challenging because predictions are commonly focused on the most discriminative part of the objects, and not necessarily the region containing the object.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

MobileNet is an efficient CNN model with depth-wise convolution layers and was adopted into many works.

“MobileNet [74, 169] was an efficient CNN model with depth-wise convolution layers which was also adopted into many works such as [170] and [171].”

• cites [74] [74] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. Andreetto, H. Adam, Mobilenets: Efficient convolutional neural networks for mobile vision applications, in: arXiv preprint arXiv:1704.04861, 2017. -Not checked yet

  Not checked in the free run.

• cites [169] [169] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, L.-C. Chen, Inverted residuals and linear bottlenecks: Mobile networks for classification, detection and segmentation, in: arXiv preprint arXiv:1801.04381, 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Though segmentation can significantly improve detection performance, obtaining pixel-level annotation is very expensive.

“Though segmentation can significantly improve detection performance, obtaining the pixel-level annotation is very expensive.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Object detection algorithms make a dense set of predictions and thus these predictions cannot be directly used for evaluation due to heavy duplication.

“Object detection algorithms make a dense set of predictions and thus these predictions cannot be directly used for evaluation due to heavy duplication.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Knowledge distillation is a training strategy which distills the knowledge in an ensemble of models into a single model via a teacher-student training scheme.

“Knowledge distillation is a training strategy which distills the knowledge in an ensemble of models into a single model via teacher-student training scheme.”

May combine two claims.

Soft-NMS avoided eliminating predictions of clustered objects and showed improvement on many common benchmarks.

“Soft-NMS avoided eliminating prediction of clustered objects and showed improvement in many common benchmarks.”

• cites [165] [165] N. Bodla, B. Singh, R. Chellappa, L. S. Davis, Soft-nms - improving object detection with one line of code, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Model compression and quantization are approaches to optimize learned models off-line.

“Another approach is to optimize models off-line, such as model compression and quantization [173, 174, 175, 176, 177, 178, 179] on the learned models.”

• cites [173] [173] Y. D. Kim, E. Park, S. Yoo, T. Choi, L. Yang, D. Shin, Compression of deep convolutional neural networks for fast and low power mobile applications, in: Computer Science, 2015. -Not checked yet

  Not checked in the free run.

• cites [174] [174] Y. He, X. Zhang, J. Sun, Channel pruning for accelerating very deep neural networks, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

• cites [175] [175] Y. Gong, L. Liu, M. Yang, L. Bourdev, Compressing deep convolutional networks using vector quantization, in: Computer Science, 2014. -Not checked yet

  Not checked in the free run.

• cites [176] [176] Y. Lin, S. Han, H. Mao, Y. Wang, W. J. Dally, Deep gradient compression: Reducing the communication bandwidth for distributed training, in: arXiv preprint arXiv:1712.01887, 2017. -Not checked yet

  Not checked in the free run.

• cites [177] [177] J. Wu, L. Cong, Y. Wang, Q. Hu, J. Cheng, Quantized convolutional neural networks for mobile devices, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [178] [178] S. Han, H. Mao, W. J. Dally, Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding, in: Fiber, 2015. -Not checked yet

  Not checked in the free run.

• cites [179] [179] S. Han, J. Pool, J. Tran, W. Dally, Learning both weights and connections for efficient neural network, in: NeurIPS, 2015. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

GBDNet extracted features from multi-scale subregions and learned a gated function to control transmission of different region information because not all contextual information is helpful for detection.

“Zeng et al. [142] proposed Gated Bi-Directional CNN (GBDNet) which extracted features from multi-scale subregions. Notably, GBDNet learned a gated function to control the transmission of different region information because not all contextual information is helpful for detection.”

• cites [142] [142] X. Zeng, W. Ouyang, B. Yang, J. Yan, X. Wang, Gated bi-directional cnn for object detection, in: ECCV, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

NVIDIA's TensorRT optimized the computation of learned models for deployment and significantly sped up inference.

“NVIDIA Corporation 1 released an acceleration toolkit TensorRT 2 which optimized the computation of learned models for deployment and thus significantly sped up the inference.”

• cites [1] [1] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [2] [2] R. Girshick, J. Donahue, T. Darrell, J. Malik, Rich feature hierarchies for accurate object detection and semantic segmentation, in: CVPR, 2014. -Not checked yet

  Not checked in the free run.

May combine two claims.

The main difficulty of training detectors from scratch is that training data of object detection is often insufficient and may lead to overfitting.

“The main difficulty of training detectors from scratch is the training data of object detection is often insufficient and may lead to overfitting.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

PVANet was proposed as a new network structure with CReLu layer to reduce nonlinear computation and accelerated inference speed.

“PVANet [104] was proposed as a new network structure with CReLu [172] layer to reduce nonlinear computation and accelerated inference speed.”

• cites [104] [104] K.-H. Kim, S. Hong, B. Roh, Y. Cheon, M. Park, Pvanet: deep but lightweight neural networks for real-time object detection, in: arXiv preprint arXiv:1608.08021, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Cascade learning is a coarse-to-fine learning strategy which collects information from the output of given classifiers to build stronger classifiers in a cascaded manner.

“Cascade learning is a coarse-to-fine learning strategy which collects information from the output of the given classifiers to build stronger classifiers in a cascaded manner.”

May combine two claims.

He et al. validated the difficulty of training detectors from scratch on MSCOCO and found that vanilla detectors could obtain competitive performance with at least 10K annotated images.

“However, later He et al. [160] validated the difficulty of training detectors from scratch on MSCOCO and found that the vanilla detectors could obtain a competitive performance with at least 10K annotated images.”

• cites [160] [160] R. G. Kaiming He, P. Dollro, Rethinking imagenet pre-training, in: arXiv preprint arXiv:1811.08883, 2018. -Not checked yet

  Not checked in the free run.

May combine two claims.

Modern object detectors heavily rely on pre-trained classification models on ImageNet, but the bias of loss functions and data distribution between classification and detection can have an adversarial impact on performance.

“Modern object detectors heavily rely on pre-trained classification models on ImageNet, however, the bias of loss functions and data distribution between classification and detection can have an adversarial impact on the performance.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Data augmentation is important for nearly all deep learning methods as they are often data-hungry and more training data leads to better results.

“Data augmentation is important for nearly all deep learning methods as they are often data-hungry and more training data leads to better results.”

May combine two claims.

Lin et al. proposed focal loss which suppressed signals from easy samples by assigning an importance weight to each sample with respect to its loss value, leading the training process to focus more on hard proposals.

“Lin et al. proposed a novel focal loss[43] which suppressed signals from easy samples. Instead of discarding all easy samples, they assigned an importance weight to each sample w.r.t its loss value as:”

• Also stated in: § 4.1.2. Imbalance Sampling · p. 21 · plain

  “The gradient signals of easy samples got suppressed which led the training process to focus more on hard proposals.”

• cites [43] [43] T.-Y. Lin, P. Goyal, R. Girshick, K. He, P. Doll´ ar, Focal loss for dense object detection, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

He et al.'s findings proved no specific structure was required for training from scratch, which contradicted previous work.

“Their findings proved no specific structure was required for training from scratch which contradicted the previous work.”

• cites [160] [160] R. G. Kaiming He, P. Dollro, Rethinking imagenet pre-training, in: arXiv preprint arXiv:1811.08883, 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

For two-stage detectors, region feature encoding is a critical step to extract features from proposals into fixed length feature vectors.

“For two-stage detectors, region feature encoding is a critical step to extract features from proposals into fixed length feature vectors.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Due to precise pixel-level annotation, segmentation feature maps capture strong spatial information.

“Due to precise pixel-level annotation, segmentation feature maps capture strong spatial information.”

May combine two claims.

Current state-of-the-art object detection algorithms can achieve very strong results on public datasets, but their inference speeds make it difficult to apply them in real applications.

“Although current state-of-the-art algorithms [167, 1] can achieve very strong results on public datasets, their inference speeds make it difficult to apply them into real applications.”

• cites [167] [167] J. Huang, V. Rathod, C. Sun, M. Zhu, A. Korattikara, A. Fathi, I. Fischer, Z. Wojna, Y. Song, S. Guadarrama, et al., Speed/accuracy trade-offs for modern convolutional object detectors, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

• cites [1] [1] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

Deformable ROI Pooling generalized aligned RoI pooling by learning an offset for each grid and adding it to the grid center; it can automatically model image content without being constrained by fixed receptive fields.

“Dai et al. proposed Deformable ROI Pooling [97] which generalized aligned RoI pooling by learning an offset for each grid and adding it to the grid center.”

• Also stated in: § 3.5.2. Region Feature Encoding · p. 18 · plain

  “Deformable ROI Pooling can automatically model the image content without being constrained by fixed receptive fields.”

• cites [97] [97] J. Dai, H. Qi, Y. Xiong, Y. Li, G. Zhang, H. Hu, Y. Wei, Deformable convolutional networks, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

In Light Head R-CNN, although computation was not shared across regions, the cost could be neglected.

“Thus, though computation was not shared across regions, but the cost could be neglected.”

• cites [168] [168] Z. Li, C. Peng, G. Yu, X. Zhang, Y. Deng, J. Sun, Light-head rcnn: In defense of two-stage object detector, in: arXiv preprint arXiv:1711.07264, 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Before the deep learning era, DPMs had been successfully used for object detection and represented objects by multiple component parts using a deformable coding method, making the detector robust to nonrigid object transformation.

“Before the deep learning era, Deformable Part based Models(DPMs) [28] had been successfully used for object detection. DPMs represented objects by multiple component parts using a deformable coding method, making the detector robust to nonrigid object transformation.”

• cites [28] [28] P. Felzenszwalb, R. Girshick, D. McAllester, D. Ramanan, Discriminatively trained mixtures of deformable part models, in: PASCAL VOC Challenge, 2008. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Soft-NMS decays the confidence score of a prediction B as a continuous function of its overlap with M instead of directly eliminating B.

“Instead of directly eliminating the prediction B , Soft-NMS decayed the confidence score of B as a continuous function F”

• cites [165] [165] N. Bodla, B. Singh, R. Chellappa, L. S. Davis, Soft-nms - improving object detection with one line of code, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

May combine two claims.

R-FCN built spatially-sensitive feature maps and extracted features with position sensitive ROI Pooling to share computation costs.

“R-FCN [52] built spatially-sensitive feature maps and extracted features with position sensitive ROI Pooling to share computation costs.”

• cites [52] [52] J. Dai, Y. Li, K. He, J. Sun, R-fcn: Object detection via region-based fully convolutional networks, in: NeurIPS, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

FPN integrated different scale features with lateral connections in a top-down fashion to build scale invariant feature maps, and multiple scale-dependent classifiers were learned on these feature pyramids.

“Lin et al. proposed Feature Pyramid Network(FPN) [39] which integrated different scale features with lateral connections in a top-down fashion to build a set of scale invariant feature maps, and multiple scale-dependent classifiers were learned on these feature pyramids.”

• cites [39] [39] T.-Y. Lin, P. Doll´ ar, R. Girshick, K. He, B. Hariharan, S. Belongie, Feature pyramid networks for object detection, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Object detection requires optimizing both localization and classification tasks, which makes it more difficult to train robust detectors than image classification.

“In contrast to image classification, object detection requires optimizing both localization and classification tasks, which makes it more difficult to train robust detectors.”

May combine two claims.

SSD combined coarse and fine features from multiple layers and made predictions from multiple layers, each responsible for a certain scale of objects.

“Liu et al.'s SSD [42] combined coarse and fine features from multiple layers together. In SSD, predictions were made from multiple layers, where each layer was responsible for a certain scale of objects.”

• cites [42] [42] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, A. C. Berg, SSD: Single shot multibox detector, in: ECCV, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

FPN showed significant improvement in object detection and other applications, and achieved state-of-the-art results in learning multi-scale features.

“FPN showed significant improvement in object detection, as well as other applications, and achieved state-of-the art results in learning multi-scale features.”

• cites [39] [39] T.-Y. Lin, P. Doll´ ar, R. Girshick, K. He, B. Hariharan, S. Belongie, Feature pyramid networks for object detection, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Non maximum suppression (NMS) is an integral part of object detection to remove duplicate false positive predictions.

“Non maximum suppression(NMS) is an integral part of object detection to remove duplicate false positive predictions”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Duplicate predictions are regarded as false positives and receive penalties in evaluation, so NMS is needed to remove these duplicate predictions.

“The duplicate predictions are regarded as false positives and will receive penalties in evaluation, so NMS is needed to remove these duplicate predictions.”

May combine two claims.

In object detection, imbalance of negative and positive samples is a critical issue; most regions of interest estimated as proposals are background images and very few are positive instances.

“In object detection, imbalance of negative and positive samples is a critical issue. That is, most of the regions of interest estimated as proposals are in fact just background images. Very few of them are positive instances (or objects).”

May combine two claims.

For object detection based on deep learning, context has not been extensively explored because convolutional networks implicitly already capture contextual information from hierarchical feature representations.

“for object detection based on deep learning, context has not been extensively explored. This is because convolutional networks implicitly already capture contextual information from hierarchical feature representations.”

Probably not a substantial claim: may not pass S3 (load-bearing).

WIDERFACE has three difficulty splits: Easy, Medium and Hard.

“According to the difficulty of detection tasks, it has three splits: Easy, Medium and Hard.”

• cites [200] [200] S. Yang, P. Luo, C.-C. Loy, X. Tang, Wwider face: A face detection benchmark, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Viola and Jones proposed a pioneering cascaded classifier using AdaBoost with Haar features for face detection and obtained excellent performance with high real-time prediction speed.

“Notably, Viola and Jones [20] proposed a pioneering cascaded classifiers using AdaBoost with Haar features for face detection and obtained excellent performance with high real time prediction speed.”

• cites [20] [20] P. Viola, M. J. Jones, Robust real-time face detection, in: IJCV, 2004. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Zhang et al. used a more extensive image pyramid structure to handle different scale objects, resizing testing images to different scales with each scale responsible for a certain scale range.

“Zhang et al. [87, 92] used a more extensive image pyramid structure to handle different scale objects. They resized the testing image to different scales and each scale was responsible for a certain scale range of objects.”

• cites [87] [87] S. Zhang, X. Zhu, Z. Lei, H. Shi, X. Wang, S. Z. Li, S3fd: Single shot scale-invariant face detector, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

• cites [92] [92] S. Zhang, L. Wen, X. Bian, Z. Lei, S. Z. Li, Single-shot refinement neural network for object detection, in: CVPR, 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Han et al. proposed Seq-NMS, which associates detection results of still images into sequences.

“Han et al. [54] proposed a Seq-NMS which associates detection results of still images into sequences.”

• cites [54] [54] W. Han, P. Khorrami, T. L. Paine, P. Ramachandran, M. Babaeizadeh, H. Shi, J. Li, S. Yan, T. S. Huang, Seq-nms for video object detection, in: arXiv preprint arXiv:1602.08465, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

Directly applying generic detection algorithms to face detection is not an optimal solution because priors can be exploited to improve face detection.

“directly applying generic detection algorithms is not an optimal solution as there could be some priors that can exploited to improve face detection.”

• AssumptionRests on an assumption: The range of scale for objects in face detection is much larger than in generic detection, and occlusion and blurred cases are more common in face detection.
• AssumptionRests on an assumption: Face objects contain strong structural information, and there is only one target category in face detection.

May combine two claims.

Horizontal Flipping was used in the testing stage and showed improvement.

“Horizontal Flipping [3, 92] was also used in the testing stage and also showed improvement.”

• cites [3] [3] K. He, G. Gkioxari, P. Doll´ ar, R. Girshick, Mask r-cnn, in: ICCV, 2017. -Not checked yet

  Not checked in the free run.

• cites [92] [92] S. Zhang, L. Wen, X. Bian, Z. Lei, S. Z. Li, Single-shot refinement neural network for object detection, in: CVPR, 2018. -Not checked yet

  Not checked in the free run.

May combine two claims.

Encoding contextual information became an effective way to improve face detection accuracy because face objects have strong physical relationships with surrounding contexts.

“Face objects have strong physical relationships with the surrounding contexts (commonly appearing with human bodies) and thus encoding contextual information became an effective way to improve detection accuracy.”

• AssumptionRests on an assumption: Face objects have strong physical relationships with surrounding contexts, commonly appearing with human bodies.

Probably not a substantial claim: may not pass S3 (load-bearing).

Wang et al. proposed a novel Repulsion Loss to detect pedestrians in a crowd.

“Wang et al. [213] proposed a novel Repulsion Loss to detect pedestrians in a crowd.”

• cites [213] [213] X. Wang, T. Xiao, Y. Jiang, S. Shao, J. Sun, C. Shen, Repulsion loss: Detecting pedestrians in a crowd, in: CVPR, 2018. -Not checked yet

  Not checked in the free run.

May combine two claims.

These learning strategies increase computation cost and were not suitable for real world applications.

“However, they also increase computation cost and thus were not suitable for real world applications.”

Probably not a substantial claim: may not pass S3 (load-bearing).

There are many other real-world applications based on object detection such as vehicle detection, traffic-sign detection and skeleton detection.

“There are also many other real-world applications based on object detection such as vehicle detection [227, 228, 229], traffic-sign detection [230, 231] and skeleton detection [232, 233].”

• cites [227] [227] Y. Zhou, L. Liu, L. Shao, M. Mellor, Dave: A unified framework for fast vehicle detection and annotation, in: ECCV, 2016. -Not checked yet

  Not checked in the free run.

• cites [228] [228] T. Gebru, J. Krause, Y. Wang, D. Chen, J. Deng, L. Fei-Fei, Fine-grained car detection for visual census estimation, in: AAAI, 2017. -Not checked yet

  Not checked in the free run.

• cites [229] [229] S. Majid Azimi, Shuffledet: Real-time vehicle detection network in onboard embedded uav imagery, in: ECCV, 2018. -Not checked yet

  Not checked in the free run.

• cites [230] [230] Z. Zhu, D. Liang, S. Zhang, X. Huang, B. Li, S. Hu, Traffic-sign detection and classification in the wild, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [231] [231] A. Pon, O. Adrienko, A. Harakeh, S. L. Waslander, A hierarchical deep architecture and mini-batch selection method for joint traffic sign and light detection, in: Conference on Computer and Robot Vision (CRV), 2018. -Not checked yet

  Not checked in the free run.

• cites [232] [232] W. Ke, J. Chen, J. Jiao, G. Zhao, Q. Ye, Srn: side-output residual network for object symmetry detection in the wild, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

• cites [233] [233] W. Shen, K. Zhao, Y. Jiang, Y. Wang, Z. Zhang, X. Bai, Object skeleton extraction in natural images by fusing scale-associated deep side outputs, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

May combine two claims.

Kang et al. proposed T-CNN, extended from Faster RCNN, which incorporated temporal and contextual information from tubelets.

“Kang et al. proposed Tubelets with Convolutional Neural Networks (T-CNN) [53] which was extended from Faster RCNN and incorporated the temporal and contextual information from tubelets(box sequence over time).”

• cites [53] [53] K. Kang, W. Ouyang, H. Li, X. Wang, Object detection from video tubelets with convolutional neural networks, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

To detect objects in videos, there are two major differences from generic detection: temporal and contextual information.

“To detect objects in videos, there are two major differences from generic detection: temporal and contextual information.”

• AssumptionRests on an assumption: The location and appearance of objects in video should be temporally consistent between adjacent frames.
• AssumptionRests on an assumption: A video consists of hundreds of frames and thus contains far richer contextual information compared to a single still image.

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

Existing detection algorithms are mainly designed for still images and are suboptimal for directly applying in videos for object detection.

“Existing detection algorithms are mainly designed for still images and are suboptimal for directly applying in videos for object detection.”

Probably not a substantial claim: may not pass S3 (load-bearing).

The annotations of WIDERFACE training and validation sets are online available.

“The annotations of training and validation sets are online available.”

• cites [200] [200] S. Yang, P. Luo, C.-C. Loy, X. Tang, Wwider face: A face detection benchmark, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

WIDERFACE has totally 32203 images with about 400k faces for a large range of scales.

“WIDERFACEhas totally 32203 images with about 400k faces for a large range of scales.”

• cites [200] [200] S. Yang, P. Luo, C.-C. Loy, X. Tang, Wwider face: A face detection benchmark, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

FDDB is a well-known benchmark with 5171 faces in 2845 images.

“The Face Detection Data set and Benchmark (FDDB) is a well-known benchmark with 5171 faces in 2845 images.”

• cites [199] [199] V. Jain, E. Learned-Miller, Fddb: A benchmark for face detection in unconstrained settings, Tech. Rep. UM-CS-2010-009, University of Massachusetts, Amherst (2010). -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Before the deep learning era, face detection was mainly based on sliding windows, with dense image grids encoded by handcrafted features followed by training classifiers to find and locate objects.

“In early stages of research before the deep learning era, face detection [20, 180, 181, 182] was mainly based on sliding windows, and dense image grids were encoded by handcrafted features followed by training classifiers to find and locate objects.”

• cites [20] [20] P. Viola, M. J. Jones, Robust real-time face detection, in: IJCV, 2004. -Not checked yet

  Not checked in the free run.

• cites [180] [180] E. Osuna, R. Freund, F. Girosit, Training support vector machines: an application to face detection, in: CVPR, 1997. -Not checked yet

  Not checked in the free run.

• cites [181] [181] M. R¨ atsch, S. Romdhani, T. Vetter, Efficient face detection by a cascaded support vector machine using haar-like features, in: Joint Pattern Recognition Symposium, 2004. -Not checked yet

  Not checked in the free run.

• cites [182] [182] S. Romdhani, P. Torr, B. Scholkopf, A. Blake, Computationally efficient face detection, in: ICCV, 2001. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

WIDERFACE has three subsets: 40% for training, 10% for validation, and 50% for test.

“It has three subsets: 40% for training, 10% for validation, and 50% for test.”

• cites [200] [200] S. Yang, P. Luo, C.-C. Loy, X. Tang, Wwider face: A face detection benchmark, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

Image pyramids are a widely used technique to improve detection results by building a hierarchical image set at different scales and making predictions on all images.

“Image pyramids [1, 92] are a widely used technique to improve detection results, which build a hierarchical image set at different scales and make predictions on all of these images.”

• cites [1] [1] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: CVPR, 2016. -Not checked yet

  Not checked in the free run.

• cites [92] [92] S. Zhang, L. Wen, X. Bian, Z. Lei, S. Z. Li, Single-shot refinement neural network for object detection, in: CVPR, 2018. -Not checked yet

  Not checked in the free run.

May combine two claims.

Deep learning based methods for pedestrian detection showed excellent performance and achieved state-of-the-art results on public benchmarks.

“Deep learning based methods for pedestrian detection [8, 9, 10, 205, 206, 207, 208, 209, 210, 211] showed excellent performance and achieved state-of-the-art results on public benchmarks.”

• cites [8] [8] J. Li, X. Liang, S. Shen, T. Xu, J. Feng, S. Yan, Scale-aware fast r-cnn for pedestrian detection, in: IEEE Transactions on Multimedia, 2018. -Not checked yet

  Not checked in the free run.

• cites [9] [9] J. Hosang, M. Omran, R. Benenson, B. Schiele, Taking a deeper look at pedestrians, in: CVPR, 2015. -Not checked yet

  Not checked in the free run.

• cites [10] [10] A. Angelova, A. Krizhevsky, V. Vanhoucke, A. S. Ogale, D. Ferguson, Real-time pedestrian detection with deep network cascades., in: BMVC, 2015. -Not checked yet

  Not checked in the free run.

• cites [205] [205] G. Brazil, X. Yin, X. Liu, Illuminating pedestrians via simultaneous detection & segmentation, in: arXiv preprint arXiv:1706.08564, 2017. -Not checked yet

  Not checked in the free run.

• cites [206] [206] X. Du, M. El-Khamy, J. Lee, L. Davis, Fused dnn: A deep neural network fusion approach to fast and robust pedestrian detection, in: IEEE Winter Conference on Applications of Computer Vision (WACV), 2017. -Not checked yet

  Not checked in the free run.

• cites [207] [207] S. Wang, J. Cheng, H. Liu, M. Tang, Pcn: Part and context information for pedestrian detection with cnns, in: arXiv preprint arXiv:1804.04483, 2018. -Not checked yet

  Not checked in the free run.

• cites [208] [208] D. Xu, W. Ouyang, E. Ricci, X. Wang, N. Sebe, Learning cross-modal deep representations for robust pedestrian detection, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

• cites [209] [209] R. Benenson, M. Omran, J. Hosang, B. Schiele, Ten years of pedestrian detection, what have we learned?, in: ECCV, 2014. -Not checked yet

  Not checked in the free run.

• cites [210] [210] Z. Cai, M. Saberian, N. Vasconcelos, Learning complexity-aware cascades for deep pedestrian detection, in: ICCV, 2015. -Not checked yet

  Not checked in the free run.

• cites [211] [211] P. Sermanet, K. Kavukcuoglu, S. Chintala, Y. LeCun, Pedestrian detection with unsupervised multi-stage feature learning, in: CVPR, 2013. -Not checked yet

  Not checked in the free run.

May combine two claims.

These learning strategies largely improved the capability of detectors to handle different scale objects and were widely used in public detection competitions.

“These learning strategies largely improved the the capability of detector to handle different scale objects and thus were widely used in public detection competitions.”

Probably not a substantial claim: may not pass S3 (load-bearing).

For CityPersons, INRIA and ETH, the log-average miss rate over 9 points ranging from 1e-2 to 100 FPPI is used to evaluate detector performance, with lower being better.

“For CityPersons, INRIA and ETH, the log-average miss rate over 9 points ranging from 1 e -2 to 100 FPPI(False Positive Per Image) is used to evaluate the performance of the detectors(lower is better).”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

CityPersons has a total of 35000 persons with an additional 13000 ignored regions, and provides both bounding box annotation of all persons and annotation of visible parts.

“A total of 35000 persons with an additional 13000 ignored regions, both bounding box annotation of all persons and annotation of visible parts are provided.”

• cites [242] [242] S. Zhang, R. Benenson, B. Schiele, Citypersons: A diverse dataset for pedestrian detection, in: CVPR, 2017. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

KITTI contains three evaluation metrics: easy, moderate and hard, with difference in the min. bounding box height, max. occlusion level, etc.

“KITTI contains three evaluation metrics: easy, moderate and hard, with difference in the min. bounding box height, max. occlusion level, etc.”

• cites [246] [246] A. Geiger, P. Lenz, C. Stiller, R. Urtasun, Vision meets robotics: The kitti dataset, 2013. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

INRIA training set has 614 positive images and 1218 negative images.

“Specifically, there are 614 positive images and 1218 negative images in the training set.”

• cites [19] [19] N. Dalal, B. Triggs, Histograms of oriented gradients for human detection, in: CVPR, 2005. -Not checked yet

  Not checked in the free run.

May combine two claims.

Although some methods have been proposed to transform anchor-based methods into anchor-free methods, there are still some limitations such as high computation cost, with large space to improve.

“Although some methods have been proposed to transform anchor-based methods into anchor-free methods (e.g. methods based on keypoints), there are still some limitations(high computation cost etc.) with large space to improve.”

Probably not a substantial claim: may not pass S3 (load-bearing).

For FDDB, true positive rate (TPR) at 1k false positives are used for evaluation.

“For FDDB, true positive rate (TPR) at 1k false positives are used for evaluation.”

Probably not a substantial claim: may not pass S3 (load-bearing).

ETH contains 1804 frames in three video clips.

“ETH [245] contains 1804 frames in three video clips”

• cites [245] [245] A. Ess, B. Leibe, L. Van Gool, Depth and appearance for mobile scene analysis, in: ICCV, 2007. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

LVIS has been proposed in order to collect richer categorical information.

“Recently, a new benchmark dataset LVIS [235] has been proposed in order to collect richer categorical information.”

• cites [235] [235] A. Gupta, P. Dollar, R. Girshick, Lvis: A dataset for large vocabulary instance segmentation, in: CVPR, 2019. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The person class in KITTI is divided into two subclasses: pedestrian and cyclist, both evaluated by mAP method.

“The person class in KITTI is divided into two subclasses: pedestrian and cyclist, both evaluated by mAP method.”

• cites [246] [246] A. Geiger, P. Lenz, C. Stiller, R. Urtasun, Vision meets robotics: The kitti dataset, 2013. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S4 (based). May combine two claims.

Contexts can contribute to or impede visual object detection results, as objects in the visual world have strong relationships and contexts are critical to better understand the visual worlds.

“Contexts can contribute or impede visual object detection results, as objects in the visual world have strong relationships, and contexts are critical to better understand the visual worlds.”

May combine two claims.

More relevant work has been proposed to address detection problems via AutoML, such as learning FPN structure and learning data augmentation policies, which show significant improvement over the baselines.

“more relevant work has been proposed to address detection problems via AutoML [249, 250], such as learning FPN structure [250] and learning data augmentation policies [251], which show significant improvement over the baselines.”

• cites [249] [249] Y. Chen, T. Yang, X. Zhang, G. Meng, C. Pan, J. Sun, Detnas: Neural architecture search on object detection, in: arXiv preprint arXiv:1903.10979, 2019. -Not checked yet

  Not checked in the free run.

• cites [250] [250] G. Ghiasi, T.-Y. Lin, Q. V. Le, Nas-fpn: Learning scalable feature pyramid architecture for object detection, in: CVPR, 2019. -Not checked yet

  Not checked in the free run.

• cites [251] [251] B. Zoph, E. D. Cubuk, G. Ghiasi, T.-Y. Lin, J. Shlens, Q. V. Le, Learning data augmentation strategies for object detection, in: arXiv preprint arXiv:1906.11172, 2019. -Not checked yet

  Not checked in the free run.

May combine two claims.

Current anchor priors are mainly manually designed, which is difficult to match multi-scale objects, and the matching strategy based on IoU is heuristic.

“Current anchor priors are mainly manually designed which is difficult to match multi-scale objects and the matching strategy based on IoU is also heuristic.”

Probably not a substantial claim: may not pass S3 (load-bearing).

Caltech comes from approximately 10 hours 30Hz VGA video recorded by a car traversing the streets in the greater Los Angeles metropolitan area.

“which comes from approximately 10 hours 30Hz VGA video recorded by a car traversing the streets in the greater Los Angeles metropolitan area.”

• cites [244] [244] P. Dollar, C. Wojek, B. Schiele, P. Perona, Pedestrian detection: An evaluation of the state of the art, in: TPAMI, 2012. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

KITTI contains 7481 labeled images of resolution 1250x375 and another 7518 images for testing.

“KITTI [246] contains 7481 labeled images of resolution 1250x375 and another 7518 images for testing.”

• cites [246] [246] A. Geiger, P. Lenz, C. Stiller, R. Urtasun, Vision meets robotics: The kitti dataset, 2013. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S3 (load-bearing).

The evaluation metric for WIDER FACE and PASCAL FACE is mean average precision (mAP) with IoU threshold as 0.5.

“the evaluation metric for WIDER FACE and PASCAL FACE is mean average precision (mAP) with IoU threshold as 0.5”

Probably not a substantial claim: may not pass S4 (based).

Currently most detectors are anchor-based methods.

“currently most detectors are anchor-based methods”

Probably not a substantial claim: may not pass S3 (load-bearing).

MSPLD addresses the low-shot detection problem in a semi-supervised learning setting where a large-scale unlabeled dataset is available.

“Multi-modal Self-Paced Learning for Detection (MSPLD) [252] addresses the low-shot detection problem in a semi-supervised learning setting where a large-scale unlabeled dataset is available.”

• cites [252] [252] X. Dong, L. Zheng, F. Ma, Y. Yang, D. Meng, Few-example object detection with model communication, in: TPAMI, 2018. -Not checked yet

  Not checked in the free run.

Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

Developing anchor-free methods has become a very hot topic in object detection.

“developing anchor-free methods becomes a very hot topic in object detection [63, 95, 240, 241, 65]”

• cites [63] [63] H. Law, J. Deng, Cornernet: Detecting objects as paired keypoints, in: ECCV, 2018. -Not checked yet

  Not checked in the free run.

• cites [95] [95] C. Zhu, Y. He, M. Savvides, Feature selective anchor-free module for single-shot object detection, in: CVPR, 2019. -Not checked yet

  Not checked in the free run.

• cites [240] [240] X. Zhou, J. Zhuo, P. Krahenbuhl, Bottom-up object detection by grouping extreme and center points, in: CVPR, 2019. -Not checked yet

  Not checked in the free run.