Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
Open accessAbstract
Object detection performance, as measured on the canonical PASCAL VOC dataset, has plateaued in the last few years. The best-performing methods are complex ensemble systems that typically combine multiple low-level image features with high-level context. In this paper, we propose a simple and scalable detection algorithm that improves mean average precision (mAP) by more than 30% relative to the previous best result on VOC 2012 -- achieving a mAP of 53.3%. Our approach combines two key insights: (1) one can apply high-capacity convolutional neural networks (CNNs) to bottom-up region proposals in order to localize and segment objects and (2) when labeled training data is scarce, supervised pre-training for an auxiliary task, followed by domain-specific fine-tuning, yields a significant performance boost. Since we combine region proposals with CNNs, we call our method R-CNN: Regions with CNN features. We also present experiments that provide insight into what the network learns, revealing a rich hierarchy of image features. Source code for the complete system is available at http://www.cs.berkeley.edu/~rbg/rcnn.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
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?
Fine-tuning the output layer and tuning all network parameters with the objective for the new task is successful on the new classes, but performance on the old classes suffers dramatically if the network is not trained on all classes jointly.
“While this framework is very successful on the new classes, its performance on the old ones suffers dramatically, if the network is not trained on all the classes jointly.”
From Incremental Learning of Object Detectors without Catastrophic Forgetting (2017) · cites this paper as [15]
In the deep learning era, bootstrap was shortly discarded in object detection during 2014-2016.
“In the deep learning era, due to the increase of computing power, bootstrap was shortly discarded in object detection during 2014-2016 [16-20].”
From Object Detection in 20 Years: A Survey (2023) · cites this paper as [16]
RCNN yields a significant performance boost on VOC07, with a large improvement of mean Average Precision (mAP) from 33.7% (DPM-v5) to 58.5%.
“RCNN yields a significant performance boost on VOC07, with a large improvement of mean Average Precision (mAP) from 33.7% (DPM-v5 [46]) to 58.5%.”
From Object Detection in 20 Years: A Survey (2023) · cites this paper as [16]
L2 loss is prevalent in early localization research.
“L2 loss is prevalent in early research [16, 20, 51]”
From Object Detection in 20 Years: A Survey (2023) · cites this paper as [16]
Girshick et al. (2014) achieve the state of the object detection results by replacing the ConvNet of Krizhevsky et al. (2012) with the authors' 16-layer model.
“For instance, Girshick et al. (2014) achieve the state of the object detection results by replacing the ConvNet of Krizhevsky et al. (2012) with our 16-layer model.”
From Very Deep Convolutional Networks for Large-Scale Image Recognition (2014) · cites this paper as Girshick et al. (2014)
R-CNN learns features unaffected by irrelevant information outside the region of interest because it is a network trained for classification on cropped image content.
“Since it is a network trained for classification on cropped image content, R-CNN learns features unaffected by irrelevant information outside the region of interest.”
From Deformable ConvNets V2: More Deformable, Better Results (2019) · cites this paper as [17]
Early patch or region-based CNN methods achieved higher accuracy than traditional methods that relied on hand-crafted features, but suffered from inefficiency for testing.
“Some early works adapted such networks for pixel labeling with patch or region-based methods [19], [20]. Such methods achieved higher accuracy than traditional methods that relied on hand-crafted features. However, they suffered from inefficiency for testing.”
From DeepIGeoS: A Deep Interactive Geodesic Framework for Medical Image Segmentation (2019) · cites this paper as [20]
Feature computation in R-CNN is time-consuming because it repeatedly applies deep convolutional networks to the raw pixels of thousands of warped regions per image.
“But the feature computation in RCNNis time-consuming, because it repeatedly applies the deep convolutional networks to the raw pixels of thousands of warped regions per image.”
From Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition (2015) · cites this paper as [7]
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.”
From Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition (2015) · cites this paper as [7]
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.”
From Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition (2015) · cites this paper as [7]
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.”
From Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition (2015) · cites this paper as [7]
RCNN achieved 31.4% on the ILSVRC2013 detection dataset.
“and 31.4% on the ILSVRC2013 detection dataset.”
From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [6]
RCNN improved upon the state-of-the-art detection algorithm by more than 30% at the time of publishing.
“At the time of publishing, it improved upon the state-of-the-art detection algorithm by more than 30%”
From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [6]
RCNN achieved up to 53.3% accuracy on the PASCAL VOC 2012 dataset.
“getting up to an accuracy of 53.3% on the PASCAL VOC 2012 dataset”
From A review on modern defect detection models using DCNNs – Deep convolutional neural networks (2022) · cites this paper as [6]
The citation lists haven’t been fetched yet
References 48 · Cited by 32,684
- Authors
- Girshick, R., Donahue, J., Darrell, T., & Malik, J.
- Affiliations
- University of California, Berkeley (US)
- Venue
- 2014 IEEE Conference on Computer Vision and Pattern Recognition
- Publisher
- IEEE
- Published
- 2014-06
- Pages
- 580-587
- ISSN
- 1063-6919 (ISSN-L)
- OpenAlex
- W2102605133
- Cited by
- 32,684
- References
- 48
- Citations by year
- 2026: 1226 · 2025: 3159 · 2024: 3864 · 2023: 4021 · 2022: 3784 · 2021: 3895 · 2020: 3398 · 2019: 3213 · 2018: 2342 · 2017: 1690 · 2016: 1147 · 2015: 752 · 2014: 177 · 2013: 8 · 2012: 2
- FWCI
- 674.95
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings - IEEE Computer Society Conference on Computer Vision and Pattern Recognition/Proceedings (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University)
- Funders
- National Science Foundation (1134072, 0905647, 1212798, MURI N000014-10-1-0933)
- Nvidia
- Defense Advanced Research Projects Agency
- Multidisciplinary University Research Initiative
- 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
- object detection, semantic segmentation, R-CNN, region proposal, convolutional neural networks, feature hierarchy, transfer learning, fine-tuning, Pascal VOC 2012, mean average precision
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.
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
