Object detection using YOLO: challenges, architectural successors, datasets and applications

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

Feature Pyramid Network is an important two-stage detector that successfully addresses some challenges by extracting features at multiple scales.

“Feature Pyramid Network (FPN) [42] is one of the important two stage object detectors, tried to overcome few of the aforementioned challenges successfully by extracting features at multiple scales of an image in the object detection process.”

• cites [42] Lin TY, Dollár P, Girshick R, He K, Hariharan B, Belongie S (2017) Feature pyramid networks for object detection. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 2117–2125 sourcePartly supported

  The Abstract (p. 1) and Introduction (p. 2) support that FPN builds multi-scale features and improves object detection when used with Faster R-CNN. They describe FPN as a generic feature extractor used in detection systems, however, rather than as a two-stage detector in its own right. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

May combine two claims.

YOLO v4 is faster and more accurate than all previous YOLO versions.

“This version is faster and more accurate than all the previous versions.”

• cites [5] Bochkovskiy A, Wang CY, Liao HY (2020) YOLOv4: optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934 sourcePartly supported

  The Introduction (p. 1) says YOLOv4 improves YOLOv3’s accuracy (AP) and speed (FPS) by 10% and 12%, respectively. The cited passages do not establish that it is faster and more accurate than every previous YOLO version. (LLM, conf=0.86)

  Experimental — accuracy not yet measured

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

Rosetta, a novel OCR system, uses Faster-RCNN to detect text characters from millions of Facebook images, while a fully convolutional CNN generates lexicon-free transcription of each word.

“A novel OCR system is developed, named as Rosetta, wherein faster-RCNN is utilized for detecting the text characters from millions of Facebook images, however, fully convolutional CNN is employed for generating the lexicon free transcription of each word [6].”

• cites [6] Borisyuk F, Gordo A, Sivakumar V (2018) Rosetta: large scale system for text detection and recognition in images. In proceedings of the 24th ACM SIGKDD international conference on knowledge discovery data mining pp 71-79 sourcePartly supported

  The Introduction (p. 1) describes Rosetta as a Facebook OCR system that uses Faster-RCNN to detect individual words and a fully convolutional CNN to produce a lexicon-free transcription of each word. The claim is inaccurate in saying Faster-RCNN detects text characters; the paper assigns character recognition to the second stage. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

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

Numerous other single-stage object detection algorithms have been introduced, including SSD, DSSD, RetinaNet, M2Det, and RefineDet++.

“There are numerous other algorithms that have been introduced in recent past such as Single Shot Detector (SSD) [43], Deconvolution Single Shot Detector (DSSD) [16], RetinaNet [41], M2Det [86], RefineDet++ [85], are based on single stage object detection.”

• cites [43] Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu CY, Berg AC (2016) Ssd: single shot multibox detector. European Conf Comput Vis 2016:21–37.  https://doi.org/10.48550/arXiv.1512.02325 sourceSupported
  Experimental — accuracy not yet measured
• cites [16] Fu CY, Liu W, Ranga A, Tyagi A, Berg AC (2017) Dssd: Deconvolutional single shot detector. arXiv preprint arXiv:1701.06659 sourceSupported
• cites [41] Lin TY, Goyal P, Girshick R, He K, Dollár P (2017) Focal loss for dense object detection. In: Proceedings of the IEEE international conference on computer vision, pp 2980–2988. https://doi.org/10.48550/arXiv.1708.02002 sourceSupported
  Experimental — accuracy not yet measured

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

GoogLeNet's inception modules contain 1×1, 3×3, and 5×5 filters to capture patterns of different sizes.

“According to [65], the inception modules contain three different sizes of filters viz. 1 × 1, 3 × 3, and 5 × 5 for capturing different sized patterns.”

• cites [65] Szegedy C, Liu W, Jia Y, Sermanet P, Reed S, Anguelov D, Erhan D, Vanhoucke V, Rabinovich A (2015) Going deeper with convolutions. In: Proceedings of the IEEE conference on computer vision and pattern recognition, pp 1–9. https://doi.org/10.48550/arXiv.1409.4842 sourceSupported

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

Global average pooling reduces the number of parameters and model complexity.

“It has resulted in a decrease in number of parameters, so as the model complexity [39].”

• cites [39] Lin M, Chen Q, Yan S (2013) Network in network. arXiv preprint arXiv:1312.4400. https://doi.org/10.48550/arXiv.1312.4400 sourceSupported

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

VGG16 and VGG19 have 16 and 19 layers, respectively.

“The authors proposed this architecture in two flavors viz. VGG19 and VGG16, consisting of 19 and 16 layers in the deep neural networks respectively [64].”

• cites [64] Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409. https://doi.org/10.48550/arXiv.1409.1556 sourceSupported

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

ResNet increases layer count to 34, nearly double VGG19.

“The number of layers is increased to 34 which is almost doubled in comparison with VGG19 [64].”

• cites [64] Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409. https://doi.org/10.48550/arXiv.1409.1556 sourceSupported

May combine two claims.

YOLO v3 is trained and evaluated on COCO and provides improved performance.

“The third version of YOLO tries to overcome aforementioned drawbacks and provides an efficient way for detecting objects with an improved performance while trained and evaluated on the COCO dataset [55].”

• cites [55] Redmon J, Farhadi A (2018) Yolov3: an incremental improvement. arXiv preprint arXiv:1804.02767 sourceSupported

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

VGG uses 3×3 filters throughout its architecture to introduce model consistency and reduce the number of parameters.

“They proposed smart factorization in the convolution operation, specifically, 3 × 3 filters are used throughout the entire architecture to introduce model consistency and to reduce the number of parameters.”

• cites [64] Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409. https://doi.org/10.48550/arXiv.1409.1556 sourcePartly supported

  The Discussion (p. 2) says the networks use small 3 × 3 receptive fields throughout and explains that stacking them uses fewer parameters than a single 7 × 7 layer. The Architecture section (p. 2) also notes a configuration with 1 × 1 filters, and the paper does not identify “model consistency” as a reason for the design. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

May combine two claims.

Object detection algorithms in deep learning require large datasets, labor-intensive annotations, and powerful computational resources.

“Object detection algorithms in deep learning need larger datasets for computation, labor intensive approaches for annotations, and powerful computational resources for processing [45].”

• cites [45] Liu L, Ouyang W, Wang X, Fieguth P, Chen J, Liu X, Pietikäinen M (2020) Deep learning for generic object detection: a survey. Int J Comput Vis 128(2):261–318. https://doi.org/10.48550/arXiv.1809.02165 sourceSupported
  Experimental — accuracy not yet measured

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

In Network-in-Network, SoftMax is the activation in the final layer and 1×1 convolution filters are placed in a multilayer perceptron after a convolutional layer, reducing dimensionality.

“Additionally, SoftMax is used as an activation function in the last layer and (1 × 1) convolution filters are placed in a multilayer perceptron after convolutional layer resulting in dimensionality reduction [23].”

• cites [23] Gu J, Wang Z, Kuen J, Ma L, Shahroudy A, Shuai B, Liu T, Wang X, Wang G, Cai J, Chen T (2018) Recent advances in convolutional neural networks. Pattern Recogn 77:354–377. https://doi.org/10.1016/j.patcog.2017.10.013 sourcePartly supported

  Section 3.1.4 (p. 5) says Network in Network places 1 × 1 convolutions after a traditional convolutional layer in its mlpconv layer and feeds the final pooled output into softmax. It does not say those convolutions reduce dimensionality; Section 3.1.5 (p. 5) assigns that role to 1 × 1 convolutions in the Inception module. (LLM, conf=0.90)

  Experimental — accuracy not yet measured

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

Real-time object detection and tracking with benchmarking results has been implemented on embedded hardware such as Jetson TX/AGX Xavier and Intel Neural Compute Stick using RCNN.

“Real time object detection and tracking is implemented and benchmarking results are obtained using embedded hardware such as Jetson TX/AGX Xavier and Intel Neural Compute Stick wherein RCNN is employed on the embedded hardware for object detection [26].”

• cites [26] Hossain S, Lee DJ (2019) Deep learning-based real-time multiple-object detection and tracking from aerial imagery via a flying robot with GPU-based embedded devices. Sensors 19(15):3371 sourcePartly supported

  The Abstract and §7 describe embedded detection and tracking with benchmark results for Jetson devices and systems using a Neural Compute Stick. Section 2.2.1 names R-CNN among the detection algorithms for the Xavier system, and §5.1 reports that Faster R-CNN was accurate but had a very low performance rate. The cited passages do not establish that R-CNN ran on the Neural Compute Stick or that its detection achieved real-time performance; §5.6 reports stick-based FPS results without identifying R-CNN. (LLM, conf=0.83)

  Experimental — accuracy not yet measured

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

ResNet is 20 times deeper than AlexNet and 8 times deeper than VGG.

“It is 20 times deeper than AlexNet and 8 times deeper than VGG [23].”

• cites [23] Gu J, Wang Z, Kuen J, Ma L, Shahroudy A, Shuai B, Liu T, Wang X, Wang G, Cai J, Chen T (2018) Recent advances in convolutional neural networks. Pattern Recogn 77:354–377. https://doi.org/10.1016/j.patcog.2017.10.013 sourceSupported

May combine two claims.

Fast R-CNN significantly improves training time and inference time over R-CNN and also increases mAP.

“Fast R-CNN [20] achieves a significant improvement in the model training time and inference time. Moreover, it also observes an increase in the performance matrix of the object detection i.e., mAP.”

• cites [20] Girshick R (2015) Fast r-cnn. In proceedings of the IEEE international conference on computer vision 2015, pp 1440–1448 sourceNot checked — no full text of the source

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

YOLO provides strong competition to both two-stage detectors and earlier single-stage detectors in accuracy and inference time.

“YOLO can be seen as giving a tough fight to not only two staged detectors but previous single staged detectors also in terms of accuracy and inference time.”

• AssumptionRests on an assumption: Single-stage object detectors generally have faster inference than two-stage detectors.

May combine two claims.

YOLO and its architectural successors have substantially improved detection accuracy, sometimes surpassing two-stage detectors.

“Moreover, with the advent of YOLO (You Only Look Once) and its architectural successors, the detection accuracy is improving significantly and sometime it is better than two stage detectors.”

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

On Pascal VOC 2007, Fast-RCNN and Faster-RCNN are 25 and 250 times faster than traditional RCNN, respectively, with similar mAP around 66.

“In the experimentational setup for object detection on Pascal VOC 2007, Fast-RCNN and Faster-RCNN are 25 and 250 times faster respectively compared to traditional RCNN with almost similar mAP of around 66.”

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

Fast YOLO is a variant with lower model complexity and 9 convolutional layers with fewer filters.

“A variant of YOLO with lesser model complexity known as Fast YOLO is proposed for faster detection of objects. It has 9 convolutional layers with comparatively lesser filters in those layers.”

May combine two claims.

A deeper ResNet became over-parameterized and produced large training error.

“Unfortunately, the network turned out to be an over parameterized network leading to large training error [25].”

• cites [25] He K, Zhang X, Ren S, Sun J (2016) Deep residual learning for image recognition. In proceedings of the IEEE conference on computer vision and pattern recognition, pp 770-778 sourceNot checked — no full text of the source

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

YOLO computes a confidence score by multiplying object probability with IoU between ground-truth and predicted box.

“Confidence score (cs) is computed for each bounding box per grid by multiplying pc with Intersection over Union (IoU) between the ground-truth and predicted-bounding-box.”

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

YOLO is not the first algorithm that uses Single Shot Detector for object detection.

“YOLO is not the first algorithm that uses Single Shot Detector (SSD) for object detection.”

May combine two claims.

YOLO v4 uses CSPDarknet53 as backbone, which has 29 convolutional layers with 3×3 filters and about 27.6 million parameters, and outperformed CSPResNeXt50 and EfficientNetB3.

“CSPDarknet53, a network with 29 convolution layers with 3 × 3 filters and approximately 27.6 million parameters, is chosen as backbone that outperformed the remaining architectures”

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

YOLO v3 uses Darknet-53 with 53 convolutional layers, 3×3 and 1×1 filters, and shortcut connections.

“This architecture is based on Darkent-53 framework, a network that uses 3 × 3 and 1 × 1 convolutional filter along with some shortcut connections, significantly larger and powerful architecture with 53 convolutional layers.”

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

YOLO is considered one of the most common choices in production because of its simple architectural design, low complexity, and easy implementation.

“It is considered as one of the most common choices in production only because of its simple architectural design, low complexity, and easy implementation.”

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

YOLO divides an image into S×S grid cells, each predicting B bounding boxes, object probability, and conditional class probabilities.

“the image/frame is divided into S × S grid cells, each grid cell predicts B bounding boxes along with their positions and dimensions, probability of an object in the underlying grid, and conditional class probabilities.”

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

ResNet152 has about 25.6 million parameters, while ResNet110 has about 1.7 million.

“In addition, ResNet152 has a total of 25.6 million parameters whereas ResNet110 has 1.7 million parameters [31].”

• cites [31] Khan A, Sohail A, Zahoora U, Qureshi AS (2020) A survey of the recent architectures of deep convolutional neural networks. Artif Intell Rev 53(8):5455–5516. https://doi.org/10.1007/s10462-018-9633-3 sourceNot checked — no full text of the source

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

YOLO v3 generates feature maps at three scales, downsampling by factors of 32, 16, and 8.

“It generates feature maps at three different scales. Specifically, it down-samples the input at 3 different scales by a factor of 32, 16, and 8.”

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

Lighter models yield faster inferences, creating a trade-off between efficacy and performance of a deployed model on TPU.

“Lighter the model, faster the inferences and vice versa, generating a trade-off between the efficacy and performance matrix of a deployed structured model on tensor processing unit (TPU).”

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

Bounding boxes are approximations of ground truth, and background pixels included in predictions affect algorithm accuracy.

“Bounding boxes are the approximations of the ground-truth. Generally, background pixels are also included during predictions, this affects the accuracy of the algorithm.”

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

YOLO v2 can operate on input dimensions from 320×320 to 608×608 after removing fully connected layers.

“After exclusion of fully connected layers, YOLO (v2) can now operate on any dimension ranging from 320 × 320 to 608 × 608.”

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

The Network-in-Network architecture stacks three mlpconv layers and then a global average pooling layer.

“Here three layers of mlpconv are stacked followed by a Global Average Pooling (GAP) layer [12].”

• cites [12] Cook A (2017) Global average pooling layers for object localization. https://alexisbcook.github.io/2017/globalaverage-poolinglayers-for-object-localization/. Accessed 19 Aug 2019Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

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

GoogLeNet has 22 layers and about 4 million parameters, which is 12 times fewer than AlexNet.

“This architecture is a deep network with 22 layers but has 12 times less parameters than AlexNet architecture, consisting of 4 million parameters approximately [31].”

• cites [31] Khan A, Sohail A, Zahoora U, Qureshi AS (2020) A survey of the recent architectures of deep convolutional neural networks. Artif Intell Rev 53(8):5455–5516. https://doi.org/10.1007/s10462-018-9633-3 sourceNot checked — no full text of the source

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

YOLO v1's architecture is inspired by GoogLeNet.

“The architecture of YOLO is inspired by the GoogLeNet architecture.”

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

YOLO v4 prefers PAN over FPN for feature aggregation.

“In addition, Path Aggregation Network (PAN) is preferred over Feature Pyramid Network (FPN) where PAN is a bottom-up path augmentation that uses Adaptive feature pooling which works similar to FPN.”

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

GoogLeNet achieved better performance than previous architectures on ImageNet.

“It records much better performance than any of the previous architectures on ImageNet dataset.”

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

YOLO v1 has 24 convolutional layers followed by 2 fully connected layers.

“In YOLO, we have 24 convolution layers followed by 2 fully connected layers as shown in Fig. 18.”

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

YOLO v2 uses Darknet-19 with 19 convolutional layers and 5 max-pooling layers.

“It uses darknet-19 framework consisting of 19 convolutional layers and 5 max pooling layers, as illustrated in Fig. 19.”

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

YOLOs perform significantly better than two-stage detectors in detection accuracy and inference time.

“YOLOs are performing significantly better in comparison with their counterpart two stage object detectors in terms of detection accuracy and inference time.”

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

YOLO uses non-max suppression based on IoU to discard overlapping bounding boxes.

“The second criteria for discarding the less relevant bounding boxes is known as non max suppression which is further based upon the IoU.”

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

ResNet uses identity mapping with skip connections of length 2.

“Subsequently, the concept of identity mapping is introduced in [11]. A skip or residual connection of length 2 is proposed as shown in Fig. 13.”

• cites [11] Choi H, Ryu S, Kim H (2018) Short-term load forecasting based on ResNet and LSTM. In IEEE international conference on communications, control, and computing Technologies for Smart Grids (SmartGridComm), pp 1-6 sourceNot checked — no full text of the source

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

VGG achieved 92.7% top-5 accuracy for object detection on the ImageNet test dataset.

“The improvised architecture is known as VGG, achieved a top-5 accuracy of 92.7% for an object detection task on the test dataset of ImageNet, a standard dataset containing around 14 million images belonging to 1 k classes.”

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

Smaller filters result in fewer parameters and can identify smaller objects more effectively.

“Smaller filters result in fewer parameters and are capable of identifying smaller objects more effectively.”

May combine two claims.

Deep learning models perform poorly on datasets with fewer instances despite outperforming traditional machine learning approaches.

“Though deep learning models outperform traditional machine learning approaches by a great margin, they demonstrate poor performance while evaluating on the datasets with fewer instances.”

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

CNN is widely adopted to capture spatial and contextual information with fewer parameters because deep neural networks have scaling inefficacy.

“Due to scaling inefficacy in deep neural networks, CNN is widely adopted to capture the spatial and contextual information with fewer parameters.”

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

Google Lens uses a region proposal network to detect character-level bounding boxes for text recognition.

“A group of researchers from Google developed an application using computer vision, machine learning, and Google’s Knowledge Graph called as Google Lens [22] wherein a region proposal network (RPN) is utilized to detect the character level bounding boxes for text recognition.”

• cites [22] Google Lens – Wikipedia (n.d.), https://en.wikipedia.org/wiki/Google_Lens. Accessed 06 Aug 2020Not matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

May combine two claims.

Convolution with anchor boxes increases recall by approximately 7% but reduces mAP by 0.3% in YOLO v2.

“An approximate increase of 7% in recall is observed by the usage of convolution with anchor boxes, however, 0.3% reduction in mAP is also noticed.”

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

YOLO v2 struggles to detect smaller objects.

“however, this version struggles in detection of smaller sized objects.”

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

VGG has approximately 138 million parameters.

“With increase in depth of network by series of convolutions, pooling layers and fully connected layers, it has resulted in 138 M parameters approximately [31].”

• cites [31] Khan A, Sohail A, Zahoora U, Qureshi AS (2020) A survey of the recent architectures of deep convolutional neural networks. Artif Intell Rev 53(8):5455–5516. https://doi.org/10.1007/s10462-018-9633-3 sourceNot checked — no full text of the source

May combine two claims.

YOLO v1 has large localization error and lower recall than two-stage detectors.

“Large localization error and lower recall in comparison with two stage object detectors may be considered as the two significant drawbacks of this version of YOLO.”

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

Model degradation means accuracy improves with depth only up to a point.

“Authors introduced the problem of model degradation i.e. model accuracy increases with an increase in the model depth to a certain extent only.”

May combine two claims.

YOLO v3 performs better for smaller objects but less accurately for medium and large objects.

“As this version outperforms for smaller sized objects, however, suffers in producing accurate results for medium and large sized objects.”

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

A YOLO grid cell predicts B×5+n values, giving an output tensor of shape S×S×(B×5+n).

“A grid cell predicts (B × 5 + n) values, where B is the number of bounding boxes per grid cell. The output tensor shape would be S × S×(B × 5 + n)”

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

GoogLeNet uses auxiliary loss with intermediate SoftMax classifiers during training to address vanishing gradients.

“In order to overcome the problem of vanishing gradients, auxiliary loss is utilized by introducing an intermediate SoftMax classifiers during model training.”

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

YOLOs make fewer background identification errors than two-stage detectors.

“Moreover, YOLOs make comparatively fewer errors in the identification of the background”

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

Combining YOLO and Fast-RCNN outperforms each standalone architecture on Pascal VOC.

“The combination of YOLO and Fast-RCNN outperform to the each of these standalone architecture on Pascal VOC dataset.”

May combine two claims.

Single-stage detectors are mainly designed for real-time object detection but lag behind two-stage detectors in performance metrics.

“Single stage object detectors are mainly designed for real time object detection; however, they are lagging in the performance metrics to the two stage object detectors.”

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

YOLO-v4 attempts to better balance accuracy and speed by introducing universal CNN optimization features.

“YOLO-v4 is also an attempt to better deal with the trade-off between the accuracy and speed of the model by introducing several universal optimization features of CNNs.”

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

Deep models take a long time to train due to large datasets and model complexity.

“Generally, deep models take huge amount of time for model training due to the large dataset size and model complexity”

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

YOLO-v2 with multi-scale training offers better detection on varying sized objects and an easy trade-off between performance and inference.

“With the help of multi scale training, YOLO-v2 offers better detection on varying sized objects with an easy trade-off between the model performance and model inferences.”

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

YOLOs are widely used in real-time object tracking such as self-driving cars.

“YOLOs are being rigorously employed in various real time object tracking such as self-driving cars.”

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

Models trained for one task may not generalize well to other similar tasks.

“Models trained for one particular task may not perform well on other similar tasks, resulting non-generalizability of the model for the data it has not seen before.”

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

YOLO-v3 is a better, fast, and accurate detector with a feature pyramid network neck.

“YOLO-v3 is better detector, fast, and accurate, necked by feature pyramid network.”