CoChat
Sign Up
Cited by 11,971Advanced Neural Network ApplicationsAdvanced Image and Video Retrieval TechniquesDomain Adaptation and Few-Shot Learning

Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition

He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, Sun, Jian·2015·IEEE Transactions on Pattern Analysis and Machine Intelligence
DOI 10.1109/tpami.2015.2389824
Open this paper in CoChat — free account

Abstract

Existing deep convolutional neural networks (CNNs) require a fixed-size (e.g., 224 × 224) input image. This requirement is "artificial" and may reduce the recognition accuracy for the images or sub-images of an arbitrary size/scale. In this work, we equip the networks with another pooling strategy, "spatial pyramid pooling", to eliminate the above requirement. The new network structure, called SPP-net, can generate a fixed-length representation regardless of image size/scale. Pyramid pooling is also robust to object deformations. With these advantages, SPP-net should in general improve all CNN-based image classification methods. On the ImageNet 2012 dataset, we demonstrate that SPP-net boosts the accuracy of a variety of CNN architectures despite their different designs. On the Pascal VOC 2007 and Caltech101 datasets, SPP-net achieves state-of-the-art classification results using a single full-image representation and no fine-tuning. The power of SPP-net is also significant in object detection. Using SPP-net, we compute the feature maps from the entire image only once, and then pool features in arbitrary regions (sub-images) to generate fixed-length representations for training the detectors. This method avoids repeatedly computing the convolutional features. In processing test images, our method is 24-102 × faster than the R-CNN method, while achieving better or comparable accuracy on Pascal VOC 2007. In ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2014, our methods rank #2 in object detection and #3 in image classification among all 38 teams. This manuscript also introduces the improvement made for this competition.

AI snapshot

Written from the paper’s full text.

TL;DR

SPP-net adds a spatial pyramid pooling layer to CNNs so they can accept arbitrary image sizes/scales and still produce fixed-length representations for classification and detection. It improves multiple CNN architectures on ImageNet, achieves strong VOC 2007/Caltech101 classification, and makes R-CNN-style detection 24–102× faster.

Key findings

  • On ImageNet 2012 validation, SPP improves four architectures: Overfeat-7 top-1 error drops from 32.01% (no SPP) to 30.36% (single-size SPP) and 29.68% (multi-size SPP); top-5 drops from 11.97% to 11.12% and 10.95%, respectively.
  • Full-image and multi-view testing help: for Overfeat-7 multi-size, 1 full view gives 31.25% top-1 vs. 32.57% for 1 crop; 96 views reduce top-5 from 10.95% to 9.36%, and adding two full-image views reduces it to 9.14%.
  • Classification: SPP-net reaches 82.44% mAP on Pascal VOC 2007 and 93.42% ± 0.5 on Caltech101 using a single full-image representation and no fine-tuning; the Caltech101 result exceeds the previous 88.54% by 4.88%.
  • Detection: SPP-net with 5-scale, fine-tuned fc7, and bounding-box regression reaches 59.2% mAP on VOC 2007, comparable to R-CNN’s 59.2% with ZF-5 and better than R-CNN’s 58.5% with AlexNet; it is 24–102× faster than R-CNN. In ILSVRC 2014, the method ranks #2 in detection (35.11% mAP after six models; 31.84% single model) and #3 in classification (8.06% top-5 after eleven models).

Methods

  • The network replaces the last pooling layer with an SPP layer on top of the last convolutional layer; SPP pools features in fixed spatial bins, e.g. a 4-level pyramid {6×6, 3×3, 2×2, 1×1} with 50 bins, producing fixed-length vectors for fully connected layers/classifiers.
  • Training uses single-size 224×224 images, or multi-size training approximated by two fixed-size networks sharing parameters (224 and 180), switching each epoch; a variant samples size uniformly from [180, 224].
  • Experiments cover ImageNet 2012 classification, Pascal VOC 2007 and Caltech101 classification, and VOC 2007/ILSVRC 2014 detection; detection uses selective search windows, SPP on feature maps, linear SVMs, fc-layer fine-tuning, and bounding-box regression.

Limitations

  • The paper has no formal limitations section; the following limitations are evident from the text.
  • Multi-size training is not truly arbitrary-size training: it is implemented as two fixed-size networks sharing parameters because GPU implementations prefer fixed input images.
  • Detection fine-tuning is limited for simplicity to the fully connected layers, and the detection experiments initially use the ZF-5 single-size trained model.
  • For ILSVRC 2014 detection, the provided-data-only setting creates a pre-training gap: DET training data is about 1/3 of CLS data, and a 200-category DET pre-trained network drops VOC mAP to 32.7% vs. 43.0% with CLS pre-training; 499-category pre-training improves it to 35.9%, and min(w,h)=400 to 37.8%, but a considerable gap remains.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/tpami.2015.2389824

Cites

54 of 54 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 56 cited sources checked · 0 supported · 0 flagged

  • Findingreasoning
    § Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition · p. 1

    The SPP-net detection method avoids repeatedly computing convolutional features.

    “This method avoids repeatedly computing the convolutional features.”
    Quote in the paperSubstantial claim
  • Findingcitation
    § Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition · p. 1

    On the Pascal VOC 2007 and Caltech101 datasets, SPP-net achieves state-of-the-art classification results using a single full-image representation and no fine-tuning.

    “On the Pascal VOC 2007 and Caltech101 datasets, SPP-net achieves state-of-theart classification results using a single full-image representation and no fine-tuning.”
    • Also stated in: § 1 INTRODUCTION · p. 2 · plain
      “SPP-net also shows state-of-the-art classification results on Caltech101 [21] and Pascal VOC 2007 [22] using only a single full-image representation and no fine-tuning.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [21] [21] L. Fei-Fei, R. Fergus, and P. Perona, 'Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories,' CVIU , 2007. -

      Not checked in the free run.

    • cites [22] [22] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman, 'The PASCAL Visual Object Classes Challenge 2007 (VOC2007) Results,' 2007. -

      Not checked in the free run.

  • Findingown data
    § Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition · p. 1

    On the ImageNet 2012 dataset, SPP-net boosts the accuracy of a variety of CNN architectures despite their different designs.

    “On the ImageNet 2012 dataset, we demonstrate that SPP-net boosts the accuracy of a variety of CNN architectures despite their different designs.”
    • Also stated in: § 1 INTRODUCTION · p. 2 · boosted
      “In a series of controlled experiments on the ImageNet 2012 dataset, we demonstrate that SPP improves four different CNN architectures in existing publications [3], [4], [5] (or their modifications), over the no-SPP counterparts.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [3] [3] A. Krizhevsky, I. Sutskever, and G. Hinton, 'Imagenet classification with deep convolutional neural networks,' in NIPS , 2012. -

      Not checked in the free run.

    • cites [4] [4] M. D. Zeiler and R. Fergus, 'Visualizing and understanding convolutional neural networks,' arXiv:1311.2901 , 2013. -

      Not checked in the free run.

    • cites [5] [5] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun, 'Overfeat: Integrated recognition, localization and detection using convolutional networks,' arXiv:1312.6229 , 2013. -

      Not checked in the free run.

  • Findingreasoning
    § Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition · p. 1

    SPP-net can generate a fixed-length representation regardless of image size or scale.

    “The new network structure, called SPP-net, can generate a fixed-length representation regardless of image size/scale.”
    • Also stated in: § 1 INTRODUCTION · p. 2 · plain
      “SPP-net not only makes it possible to generate representations from arbitrarily sized images/windows for testing, but also allows us to feed images with varying sizes or scales during training.”
    Quote in the paperSubstantial claim
  • Findingcitation
    § 1 INTRODUCTION · p. 2

    With EdgeBoxes, the SPP-net system takes 0.5 seconds to process an image including all steps.

    “With the recent fast proposal method of EdgeBoxes [25], our system takes 0.5 seconds processing an image ( including all steps ).”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [25] [25] C. L. Zitnick and P. Doll´ ar, 'Edge boxes: Locating object proposals from edges,' in ECCV , 2014. -

      Not checked in the free run.

  • May combine two claims. 45
  • Classified as probably not a substantial claim. 23
  • The quote only partly states this claim. 5

  • May combine two claims.

    Findingnone given
    § Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition · p. 1

    Using SPP-net, feature maps are computed from the entire image only once, and features are then pooled in arbitrary regions to generate fixed-length representations for training detectors.

    “Using SPP-net, we compute the feature maps from the entire image only once, and then pool features in arbitrary regions (sub-images) to generate fixed-length representations for training the detectors.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition · p. 1

    In ILSVRC 2014, the authors' methods rank #2 in object detection and #3 in image classification among all 38 teams.

    “In ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2014, our methods rank #2 in object detection and #3 in image classification among all 38 teams.”
    • Also stated in: § 1 INTRODUCTION · p. 2 · plain
      “Based on this work, we attended the competition of ILSVRC 2014 [26], and ranked #2 in object detection and #3 in image classification (both are provided-data-only tracks) among all 38 teams.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [26] [26] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al. , 'Imagenet large scale visual recognition challenge,' arXiv:1409.0575 , 2014. -

      Not checked in the free run.

  • May combine two claims.

    Findingreasoning
    § Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition · p. 1

    In processing test images, the SPP-net method is 24-102× faster than the R-CNN method while achieving better or comparable accuracy on Pascal VOC 2007.

    “In processing test images, our method is 24-102 × faster than the R-CNN method, while achieving better or comparable accuracy on Pascal VOC 2007.”
    • Also stated in: § 1 INTRODUCTION · p. 2 · plain
      “In our experiment, the SPP-net-based system (built upon the R-CNN pipeline) computes features 24-102 × faster than R-CNN, while has better or comparable accuracy.”
    Quote in the paperSubstantial claim

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

    Findingown data
    § 2.1 Convolutional Layers and Feature Maps · p. 3

    Some conv5 filters are activated by specific semantic shapes: the 55th filter by a circle shape, the 66th by a ∧-shape, and the 118th by a ∨-shape.

    “For example, the 55-th filter (Figure 2, bottom left) is most activated by a circle shape; the 66-th filter (Figure 2, top right) is most activated by a ∧ -shape; and the 118-th filter (Figure 2, bottom right) is most activated by a ∨ -shape.”
    Quote in the paperNot a substantial claim

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

    Findingnone given
    § 2.1 Convolutional Layers and Feature Maps · p. 3

    The feature maps in Figure 2 were generated without fixing the input size.

    “It is worth noticing that we generate the feature maps in Figure 2 without fixing the input size.”
    Quote in the paperNot a substantial claim

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

    Findingnone given
    § 3.1 Experiments on ImageNet 2012 Classification · p. 5

    All networks in the paper can be trained on a single GeForce GTX Titan GPU with 6 GB memory within two to four weeks.

    “All networks in this paper can be trained on a single GeForce GTX Titan GPU (6 GB memory) within two to four weeks.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 1 INTRODUCTION · p. 2

    Multi-size training converges just as traditional single-size training and leads to better testing accuracy.

    “Experiments show that this multi-size training converges just as the traditional single-size training, and leads to better testing accuracy.”
    • Also stated in: § Multi-size training · p. 4 · plain
      “In experiments, we find the convergence rate of this multi-size training to be similar to the above single-size training.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.1.2 Multi-level Pooling Improves Accuracy · p. 5

    The 30-bin pyramid result is similar to the 50-bin pyramid result (34.98/14.14) but considerably better than the no-SPP counterpart (35.99/14.76).

    “This result is similar to the 50-bin pyramid above (34.98/14.14), but considerably better than the no-SPP counterpart (35.99/14.76).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 3.1.3 Multi-size Training Improves Accuracy · p. 6

    The top-1 error of SPP-net (Overfeat-7) drops to 29.68%, which is 2.33% better than its no-SPP counterpart and 0.68% better than its single-size trained counterpart.

    “The top-1 error of SPP-net (Overfeat-7) drops to 29.68%, which is 2.33% better than its noSPP counterpart and 0.68% better than its single-size trained counterpart.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data
    § 3.1.3 Multi-size Training Improves Accuracy · p. 6

    With a random size uniformly sampled from [180, 224], SPP-net (Overfeat-7) has top-1/top-5 error of 30.06%/10.96%.

    “The top-1/5 error of SPP-net (Overfeat-7) is 30.06%/10.96%.”
    Quote in the paperNot a substantial claim

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

    Findingcitation
    § 3.1.1 Baseline Network Architectures · p. 5

    The authors' replication of ZF-5 is better than the one reported in Zeiler and Fergus.

    “Our replication of ZF-5 is better than the one reported in [4].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [4] [4] M. D. Zeiler and R. Fergus, 'Visualizing and understanding convolutional neural networks,' arXiv:1311.2901 , 2013. -

      Not checked in the free run.

  • May combine two claims.

    Findingnone given
    § 2.2 The Spatial Pyramid Pooling Layer · p. 3

    With spatial pyramid pooling, the input image can be of any size, allowing arbitrary aspect ratios and arbitrary scales.

    “With spatial pyramid pooling, the input image can be of any sizes. This not only allows arbitrary aspect ratios, but also allows arbitrary scales.”
    Quote in the paperSubstantial claim

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

    Findingreasoning
    § 4.1 Detection Algorithm · p. 10

    Because we only fine-tune the fc layers, training is very fast and takes about 2 hours on the GPU (excluding pre-caching feature maps which takes about 1 hour).

    “Because we only fine-tune the fc layers, the training is very fast and takes about 2 hours on the GPU (excluding pre-caching feature maps which takes about 1 hour).”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingreasoning
    § 3.2 Experiments on VOC 2007 Classification · p. 8

    Replacing the no-SPP net with SPP-net on the center 224x224 crop improves the fc layer results.

    “The results of the fc layers improve.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.4 Model Combination for Detection · p. 11

    After combination, mAP is boosted to 60.9%; in 17 of 20 categories the combination performs better than either individual model.

    “After combination, the mAP is boosted to 60.9% (Table 12). In 17 out of all 20 categories the combination performs better than either individual model.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.2 Detection Results · p. 10

    After bounding box regression, our 5-scale result (59.2%) is 0.7% better than R-CNN (58.5%), and our 1-scale result (58.0%) is 0.5% worse.

    “After bounding box regression, our 5-scale result ( 59.2 %) is 0.7% better than R-CNN (58.5%), and our 1-scale result (58.0%) is 0.5% worse.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given
    § 4.3 Complexity and Running Time · p. 10

    The complexity of convolutional feature computation in R-CNN is O(n·227^2) with n~2000; our method is O(r·s^2) at scale s, where r is aspect ratio.

    “The complexity of the convolutional feature computation in R-CNN is O ( n · 227 2 ) with the window number n ( ∼ 2000). This complexity of our method is O ( r · s 2 ) at a scale s , where r is the aspect ratio.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingreasoning
    § 3.3 Experiments on Caltech101 · p. 8

    In Caltech101, the fully-connected layers are less accurate (than in Pascal VOC).

    “The fully-connected layers are less accurate, and the SPP layers are better.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 4.4 Model Combination for Detection · p. 11

    The second network's mAP is comparable with the first (59.1% vs 59.2%) and outperforms the first in 11 categories.

    “Its mAP is comparable with the first network (59.1% vs . 59.2%), and outperforms the first network in 11 categories.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Complexity and Running Time · p. 11

    Considering both convolutional and fully-connected features, our 1-scale version is 102× faster than R-CNN and is 1.2% inferior; our 5-scale version is 38× faster and has comparable results.

    “Considering both convolutional and fully-connected features, our 1-scale version is 102 × faster than R-CNN and is 1.2% inferior; our 5-scale version is 38 × faster and has comparable results.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.4 Model Combination for Detection · p. 11

    The complementarity is mainly because of the convolutional layers; combining two randomly initialized fine-tuned results of the same convolutional model gave no gain.

    “We further find that the complementarity is mainly because of the convolutional layers. We have tried to combine two randomly initialized fine-tuned results of the same convolutional model, and found no gain.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data
    § 3.3 Experiments on Caltech101 · p. 8

    On the SPP (ZF-5) model, warping to 224x224 gives accuracy 89.91% using SPP layer features, lower than 91.44% on the undistorted full image.

    “On the SPP (ZF-5) model, the accuracy is 89.91% using the SPP layer as features - lower than 91.44% which uses the same model on the undistorted full image.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Findingreasoning
    § 3.3 Experiments on Caltech101 · p. 8

    In Caltech101, the SPP layers are better (than in Pascal VOC).

    “The fully-connected layers are less accurate, and the SPP layers are better.”
    Quote in the paperNot a substantial claim

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

    Findingnone given
    § 3.2 Experiments on VOC 2007 Classification · p. 7

    The method can generate a full-view image representation.

    “Our method can generate a full-view image representation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning
    § 4.3 Complexity and Running Time · p. 10

    Assuming r is about 4/3, single-scale s=688 complexity is about 1/160 of R-CNN's; 5-scale is about 1/24.

    “Assume r is about 4/3. In the single-scale version when s = 688 , this complexity is about 1/160 of R-CNN's; in the 5-scale version, this complexity is about 1/24 of R-CNN's.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The aspect ratio r is about 4/3.

  • May combine two claims.

    Findingown data
    § 4.3 Complexity and Running Time · p. 11

    Using both SS and EdgeBox in training and only EdgeBoxes in testing, mAP is 56.3 without bounding box regression, better than 55.2 due to additional training samples.

    “The mAP is 56.3 without bounding box regression, which is better than 55.2 (Table 10) due to additional training samples.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Complexity and Running Time · p. 11

    R-CNN takes 14.37s per image for convolutions, our 1-scale takes 0.053s, so ours is 270× faster; 5-scale takes 0.293s, so 49× faster.

    “R-CNN takes 14.37s per image for convolutions, while our 1-scale version takes only 0.053s per image. So ours is 270 × faster than R-CNN. Our 5-scale version takes 0.293s per image for convolutions, so is 49 × faster than R-CNN.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data
    § 4.3 Complexity and Running Time · p. 11

    Our method is 24× to 64× faster than R-CNN using AlexNet.

    “Our method is 24 × to 64 × faster.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.3 Complexity and Running Time · p. 10

    Despite having comparable accuracy, our method is much faster than R-CNN.

    “Despite having comparable accuracy, our method is much faster than R-CNN.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data
    § 4.5 ILSVRC 2014 Detection · p. 12

    Our system shows great advantages on speed for this dataset; single model takes 0.6s per testing image on GPU extracting conv features from all 5 scales; RCNN takes 32s per image; for 40k testing images, our method requires 8 GPU·hours vs RCNN 15 GPU·days.

    “It takes our single model 0.6 seconds (0.5 for conv, 0.1 for fc, excluding proposals) per testing image on a GPU extracting convolutional features from all 5 scales. Using the same model, it takes 32 seconds per image in the way of RCNN. For the 40k testing images, our method requires 8 GPU · hours to compute convolutional features, while RCNN would require 15 GPU · days.”
    Quote in the paperSubstantial claim

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

    Premisecitation
    § 1 INTRODUCTION · p. 2

    Training or running a detector on feature maps rather than image regions is a more popular idea.

    “Note that training/running a detector on the feature maps (rather than image regions) is actually a more popular idea [23], [24], [20], [5].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [23] [23] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan, 'Object detection with discriminatively trained partbased models,' PAMI , 2010. - source

      Not checked in the free run.

    • cites [24] [24] N. Dalal and B. Triggs, 'Histograms of oriented gradients for human detection,' in CVPR , 2005. - source

      Not checked in the free run.

    • cites [20] [20] K. E. van de Sande, J. R. Uijlings, T. Gevers, and A. W. Smeulders, 'Segmentation as selective search for object recognition,' in ICCV , 2011. - source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation
    § 1 INTRODUCTION · p. 2

    In R-CNN, the features from candidate windows are extracted via deep convolutional networks.

    “In the leading object detection method R-CNN [7], the features from candidate windows are extracted via deep convolutional networks.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7] [7] R. Girshick, J. Donahue, T. Darrell, and J. Malik, 'Rich feature hierarchies for accurate object detection and semantic segmentation,' in CVPR , 2014. - source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 1 INTRODUCTION · p. 1

    The cropped region may not contain the entire object, while warped content may result in unwanted geometric distortion.

    “But the cropped region may not contain the entire object, while the warped content may result in unwanted geometric distortion.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given
    § 1 INTRODUCTION · p. 2

    SPP uses multi-level spatial bins, while sliding window pooling uses only a single window size.

    “2) SPP uses multi-level spatial bins, while the sliding window pooling uses only a single window size.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 1 INTRODUCTION · p. 2

    SPP is able to generate a fixed-length output regardless of the input size, while the sliding window pooling used in previous deep networks cannot.

    “1) SPP is able to generate a fixedlength output regardless of the input size, while the sliding window pooling used in the previous deep networks [3] cannot;”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3] [3] A. Krizhevsky, I. Sutskever, and G. Hinton, 'Imagenet classification with deep convolutional neural networks,' in NIPS , 2012. -

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisenone given
    § 2.3 Training the Network · p. 4

    GPU implementations such as cuda-convnet and Caffe are preferably run on fixed input images.

    “But in practice the GPU implementations (such as cuda-convnet [3] and Caffe [35]) are preferably run on fixed input images.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [3] [3] A. Krizhevsky, I. Sutskever, and G. Hinton, 'Imagenet classification with deep convolutional neural networks,' in NIPS , 2012. -

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

    • cites [35] [35] Y. Jia, 'Caffe: An open source convolutional architecture for fast feature embedding,' http://caffe.berkeleyvision.org/, 2013. -

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

  • May combine two claims.

    Interpretationreasoning
    § 1 INTRODUCTION · p. 2

    SPP-net inherits the power of deep CNN feature maps and the flexibility of SPP on arbitrary window sizes, leading to outstanding accuracy and efficiency.

    “But SPP-net inherits the power of the deep CNN feature maps and also the flexibility of SPP on arbitrary window sizes, which leads to outstanding accuracy and efficiency.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 3.1.3 Multi-size Training Improves Accuracy · p. 6

    Previous CNN solutions deal with various scales or sizes, but they are mostly based on testing.

    “There are previous CNN solutions [5], [36] that deal with various scales/sizes, but they are mostly based on testing.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [5] [5] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun, 'Overfeat: Integrated recognition, localization and detection using convolutional networks,' arXiv:1312.6229 , 2013. - source

      Not checked in the free run.

    • cites [36] [36] A. G. Howard, 'Some improvements on deep convolutional neural network based image classification,' ArXiv:1312.5402 , 2013. - source

      Not checked in the free run.

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

    Interpretationreasoning
    § 3.1.3 Multi-size Training Improves Accuracy · p. 6

    The top-1 error of the random-size version is slightly worse than the two-size version, possibly because the size 224 used for testing is visited less.

    “The top1 error is slightly worse than the two-size version, possibly because the size of 224 (which is used for testing) is visited less.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation
    § 3.1.1 Baseline Network Architectures · p. 5

    The gain in the ZF-5 replication is because the corner crops are from the entire image, as also reported by Howard.

    “This gain is because the corner crops are from the entire image, as is also reported in [36].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36] [36] A. G. Howard, 'Some improvements on deep convolutional neural network based image classification,' ArXiv:1312.5402 , 2013. - source

      Not checked in the free run.

  • May combine two claims.

    Interpretationnone given
    § 1 INTRODUCTION · p. 2

    Training with variable-size images increases scale-invariance and reduces over-fitting.

    “Training with variable-size images increases scale-invariance and reduces over-fitting.”
    Quote in the paperNot a substantial claim

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

    Premisecitation
    § 1 INTRODUCTION · p. 2

    R-CNN shows remarkable detection accuracy on both the VOC and ImageNet datasets.

    “This method shows remarkable detection accuracy on both the VOC and ImageNet datasets.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [7] [7] R. Girshick, J. Donahue, T. Darrell, and J. Malik, 'Rich feature hierarchies for accurate object detection and semantic segmentation,' in CVPR , 2014. - source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning
    § 1 INTRODUCTION · p. 1

    Convolutional layers do not require a fixed image size and can generate feature maps of any sizes.

    “In fact, convolutional layers do not require a fixed image size and can generate feature maps of any sizes.”
    • Also stated in: § 2.1 Convolutional Layers and Feature Maps · p. 3 · plain
      “On the other hand, the convolutional layers accept inputs of arbitrary sizes.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationcitation
    § 3.1.2 Multi-level Pooling Improves Accuracy · p. 5

    The gain of multi-level pooling is not simply due to more parameters; rather, multi-level pooling is robust to variance in object deformations and spatial layout.

    “It is worth noticing that the gain of multi-level pooling is not simply due to more parameters; rather, it is because the multi-level pooling is robust to the variance in object deformations and spatial layout [15].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [15] [15] S. Lazebnik, C. Schmid, and J. Ponce, 'Beyond bags of features: Spatial pyramid matching for recognizing natural scene categories,' in CVPR , 2006. -

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisenone given
    § 1 INTRODUCTION · p. 1

    Fully-connected layers need to have fixed-size or fixed-length input by their definition.

    “On the other hand, the fully-connected layers need to have fixedsize/length input by their definition.”
    • Also stated in: § 2.1 Convolutional Layers and Feature Maps · p. 3 · plain
      “the requirement of fixed sizes is only due to the fully-connected layers that demand fixed-length vectors as inputs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning
    § 2.2 The Spatial Pyramid Pooling Layer · p. 3

    Spatial bins have sizes proportional to the image size, so the number of bins is fixed regardless of the image size.

    “These spatial bins have sizes proportional to the image size, so the number of bins is fixed regardless of the image size.”
    Quote in the paperSubstantial claim

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

    Interpretationnone given
    § 4.1 Detection Algorithm · p. 9

    Our method can be improved by multi-scale feature extraction.

    “Our method can be improved by multi-scale feature extraction.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning
    § 3.3 Experiments on Caltech101 · p. 8

    The difference in Caltech101 results is possibly because object categories are less related to those in ImageNet and deeper layers are more category-specialized.

    “This is possibly because the object categories in Caltech101 are less related to those in ImageNet, and the deeper layers are more category-specialized.”
    Quote in the paperNot a substantial claim

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

    Interpretationreasoning
    § 3.3 Experiments on Caltech101 · p. 8

    The best scale 224 on Caltech101 is mainly because objects occupy large regions, as in ImageNet.

    “This is mainly because the objects in Caltech101 also occupy large regions of the images, as is the case of ImageNet.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § 4.2 Detection Results · p. 10

    The R-CNN result is boosted by this pre-trained model because of the better architecture of ZF-5 than AlexNet and the multi-level pooling of SPPnet.

    “This is because of the better architecture of ZF-5 than AlexNet, and also because of the multi-level pooling of SPPnet (if using the noSPP ZF-5, the R-CNN result drops).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § 3.2 Experiments on VOC 2007 Classification · p. 8

    The best s = 392 is mainly because objects occupy smaller regions in VOC 2007 but larger regions in ImageNet, so relative object scales differ between the two sets.

    “This is mainly because the objects occupy smaller regions in VOC 2007 but larger regions in ImageNet, so the relative object scales are different between the two sets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § 3.1.3 Multi-size Training Improves Accuracy · p. 6

    In Overfeat and Howard's method, the single network is applied at multiple scales in the testing stage and the scores are averaged.

    “In Overfeat [5] and Howard's method [36], the single network is applied at multiple scales in the testing stage, and the scores are averaged.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [5] [5] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun, 'Overfeat: Integrated recognition, localization and detection using convolutional networks,' arXiv:1312.6229 , 2013. - source

      Not checked in the free run.

    • cites [36] [36] A. G. Howard, 'Some improvements on deep convolutional neural network based image classification,' ArXiv:1312.5402 , 2013. - source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning
    § 3.1.6 Summary and Results for ILSVRC 2014 · p. 7

    Since the advantages of SPP-net are independent of architectures, SPP-net is expected to further improve deeper and larger convolutional architectures.

    “Since the advantages of SPPnet should be in general independent of architectures, we expect that it will further improve the deeper and larger convolutional architectures [33], [32].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [33] [33] K. Simonyan and A. Zisserman, 'Very deep convolutional networks for large-scale image recognition,' arXiv:1409.1556 , 2014. - source

      Not checked in the free run.

    • cites [32] [32] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, 'Going deeper with convolutions,' arXiv:1409.4842 , 2014. - source

      Not checked in the free run.

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

    Premisecitation
    § 3.1.5 Multi-view Testing on Feature Maps · p. 7

    Overfeat's views cannot have arbitrary sizes; rather, the windows are those where the pooled features match the desired dimensionality.

    “However, their views cannot have arbitrary sizes; rather, the windows are those where the pooled features match the desired dimensionality.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [5] [5] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun, 'Overfeat: Integrated recognition, localization and detection using convolutional networks,' arXiv:1312.6229 , 2013. -

      Not checked in the free run.

  • May combine two claims.

    Premisecitation
    § 4 SPP-NET FOR OBJECT DETECTION · p. 9

    R-CNN repeatedly applies the deep convolutional network to about 2,000 windows per image, making it time-consuming; feature extraction is the major timing bottleneck in testing.

    “because R-CNN repeatedly applies the deep convolutional network to about 2,000 windows per image, it is time-consuming. Feature extraction is the major timing bottleneck in testing.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7] [7] R. Girshick, J. Donahue, T. Darrell, and J. Malik, 'Rich feature hierarchies for accurate object detection and semantic segmentation,' in CVPR , 2014. - source

      Not checked in the free run.

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

    Premisecitation
    § 3.1.3 Multi-size Training Improves Accuracy · p. 6

    Howard further trains two different networks on low/high-resolution image regions and averages the scores.

    “Howard further trains two different networks on low/highresolution image regions and averages the scores.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [36] [36] A. G. Howard, 'Some improvements on deep convolutional neural network based image classification,' ArXiv:1312.5402 , 2013. - source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § 4.2 Detection Results · p. 10

    The inferiority of non-fine-tuned fc6 is explained by fc layers pretrained on image regions but used on feature map regions, which can have strong activations near window boundaries; fine-tuning addresses this.

    “An explanation is that our fc layers are pretrained using image regions, while in the detection case they are used on the feature map regions. The feature map regions can have strong activations near the window boundaries, while the image regions may not. This difference of usages can be addressed by fine-tuning.”
    Quote in the paperSubstantial claim

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

    Premisecitation
    § 3.1.5 Multi-view Testing on Feature Maps · p. 7

    In the Overfeat paper, the views are also extracted from the convolutional feature maps instead of image crops.

    “In the Overfeat paper [5], the views are also extracted from the convolutional feature maps instead of image crops.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [5] [5] P. Sermanet, D. Eigen, X. Zhang, M. Mathieu, R. Fergus, and Y. LeCun, 'Overfeat: Integrated recognition, localization and detection using convolutional networks,' arXiv:1312.6229 , 2013. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationown data
    § 4.3 Complexity and Running Time · p. 11

    This makes our method practical for real-world applications.

    “This makes our method practical for real-world applications.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data
    § 3.2 Experiments on VOC 2007 Classification · p. 8

    These results indicate scale matters in classification tasks, and SPP-net can partially address the scale mismatch issue.

    “These results indicate scale matters in the classification tasks, and SPP-net can partially address this 'scale mismatch' issue.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationnone given
    § 5 CONCLUSION · p. 12

    The resulting SPP-net shows outstanding accuracy in classification/detection tasks and greatly accelerates DNN-based detection.

    “The resulting SPP-net shows outstanding accuracy in classification/detection tasks and greatly accelerates DNN-based detection.”
    Quote in the paperNot a substantial claim

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

    Premisecitation
    § 4.5 ILSVRC 2014 Detection · p. 12

    The winning result is 37.21% from NUS, which uses contextual information.

    “The winning result is 37.21% from NUS, which uses contextual information.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [26] [26] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al. , 'Imagenet large scale visual recognition challenge,' arXiv:1409.0575 , 2014. -

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given
    § 5 CONCLUSION · p. 12

    Many time-proven techniques/insights in computer vision can still play important roles in deep-networks-based recognition.

    “Our studies also show that many time-proven techniques/insights in computer vision can still play important roles in deep-networks-based recognition.”
    Quote in the paperNot a substantial claim

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

    Premisecitation
    § 4.3 Complexity and Running Time · p. 11

    AlexNet is faster because it uses splitting on some layers, which was designed for two GPUs in [3].

    “The AlexNet is faster because it uses splitting on some layers, which was designed for two GPUs in [3].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [3] [3] A. Krizhevsky, I. Sutskever, and G. Hinton, 'Imagenet classification with deep convolutional neural networks,' in NIPS , 2012. -

      Not checked in the free run.

  • May combine two claims.

    Interpretationown data
    § 4.5 ILSVRC 2014 Detection · p. 12

    There is still a considerable gap to the CLS-pre-training result, indicating the importance of big data to deep learning.

    “Even so, we see that there is still a considerable gap to the CLS-pre-training result. This indicates the importance of big data to deep learning.”
    Quote in the paperNot a substantial claim
65 of 65

  • The paper relies on this without support.

    SPP has not been considered in the context of CNNs.

    “Nevertheless, SPP has not been considered in the context of CNNs.”

  • The paper relies on this without support.

    The authors' method is the first one that trains a single network with input images of multiple sizes.

    “To our knowledge, our method is the first one that trains a single network with input images of multiple sizes.”

  • The authors state this as an assumption.

    The aspect ratio r is about 4/3.

    “Assume r is about 4/3.”

  • The paper relies on this without support.

    The distributions of object scales are different between DET and CLS training sets; the dominant object scale in CLS is about 0.8 of the image length, and in DET about 0.5.

    “Third, the distributions of object scales are different between DET/CLS training sets. The dominant object scale in CLS is about 0.8 of the image length, but in DET is about 0.5.”

  • The paper relies on this without support.

    The usage of the network does not depend on scale.

    “Because the usage of our network does not depend on scale, we resize the images so that the smaller dimension is s and use the same network to extract features.”
5 of 5

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Using feature pyramids, PANet, and SPP blocks improves object identification on unseen data.

    “Using feature pyramids, PANet, and SPP blocks improves object identification on unseen data as well.”

    From Real-Time Vehicle Classification and Tracking Using a Transfer Learning-Improved Deep Learning Network (2022) · cites this paper as [55]

    Section 3.1.2 (p. 5) reports that spatial pyramid pooling improved accuracy over no-SPP baselines on the ImageNet validation set. The paper does not evaluate PANet or show that the stated combination of feature pyramids, PANet, and SPP blocks improves identification on unseen data. (LLM, conf=0.90)

    Experimental — accuracy not yet measured

  • The CSP and PANet neck with an SPP block generates feature pyramids that help the overall network generalize objects of different scales and sizes.

    “This neck generates feature pyramids that help the overall network to generalize objects of different scales and sizes.”

    From Real-Time Vehicle Classification and Tracking Using a Transfer Learning-Improved Deep Learning Network (2022) · cites this paper as [55]

    The Conclusion (p. 12) describes SPP as a solution for handling different scales and sizes, and the Introduction (p. 2) says it pools features at multiple levels. The paper does not establish that a CSP/PANet neck generates feature pyramids or that this particular neck improves generalization across object scales. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • SPPNet achieved 60.9% mAP on the PASCAL VOC 2007 dataset (best one).

    “The accuracy obtained on the PASCAL VOC 2007 dataset was 60.9% mAP (best one)”

    From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [79]

  • SPPNet achieved 35.11% mAP on ISLVRC 2014, coming in second.

    “while on ISLVRC 2014 it obtained 35.11% mAP, coming in second.”

    From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [79]

  • In medical imaging, the most commonly employed deep learning methods are convolutional neural networks (CNNs).

    “In medical imaging, the most commonly employed deep learning methods are convolutional neural networks (CNNs) [1–8].”

    From Deep Learning Strategies for Ultrasound in Pregnancy (2020) · cites this paper as [8]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Stated pooling rule does not give equal-length outputs at both training sizes

    Major

    For the stated 6 × 6 pyramid level, a 10 × 10 feature map gives a pooling window of 2 and stride of 1 under this rule, producing 9 × 9 outputs rather than 6 × 6. The 13 × 13 map produces 6 × 6, so the two networks cannot have the same output length as claimed under the stated implementation.

  • Reported GPU-hours exceed the stated convolution time

    Minor

    At the stated 0.5 seconds of convolution per image, 40k images require about 5.6 GPU-hours, not 8 GPU-hours. Even the stated 0.6-second total per image amounts to about 6.7 GPU-hours.

2 of 2

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
He, K., Zhang, X., Ren, S., & Sun, J.
Affiliations
  • Xi'an Jiaotong University (CN)
  • University of Science and Technology of China (CN)
ORCID
Venue
IEEE Transactions on Pattern Analysis and Machine Intelligence
Publisher
Institute of Electrical and Electronics Engineers (IEEE)
Published
2015-09-01
Volume
37
Issue
9
Pages
1904-1916
ISSN
  • 0162-8828 (ISSN-L)
  • 2160-9292
DOI
10.1109/tpami.2015.2389824
OpenAlex
W2109255472
Cited by
11,971
References
54
Citations by year
2026: 538 · 2025: 1321 · 2024: 1718 · 2023: 1892 · 2022: 1748 · 2021: 1438 · 2020: 1062 · 2019: 874 · 2018: 611 · 2017: 447 · 2016: 219 · 2015: 93 · 2014: 7 · 2012: 2
FWCI
395.62
Type
journal article
Language
en
Open access
closed
Licence
ieeexplore.ieee.org/Xplorehelp/downloads/license-information/IEEE.html (vor)
Locations
  • IEEE Transactions on Pattern Analysis and Machine Intelligence (publishedVersion)
  • PubMed (publishedVersion)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
spatial pyramid pooling, deep convolutional neural networks, image classification, object detection, scale invariance, feature reuse, ImageNet, Pascal VOC 2007
MeSH terms
Algorithms, Animals, Humans, Image Processing, Computer-Assisted, Pattern Recognition, Automated, Neural Networks, Computer

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review