Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 52 of the 52 references it cites were found in a registry or an index; 3 of the 20 sources its claims cite were checked for support.

Object Detection Networks on Convolutional Feature Maps

Open access
Ren, Shaoqing, He, Kaiming, Girshick, Ross, Zhang, Xiangyu, Sun, Jian · 2017 · IEEE Transactions on Pattern Analysis and Machine Intelligence

Abstract

Most object detectors contain two important components: a feature extractor and an object classifier. The feature extractor has rapidly evolved with significant research efforts leading to better deep convolutional architectures. The object classifier, however, has not received much attention and many recent systems (like SPPnet and Fast/Faster R-CNN) use simple multi-layer perceptrons. This paper demonstrates that carefully designing deep networks for object classification is just as important. We experiment with region-wise classifier networks that use shared, region-independent convolutional features. We call them "Networks on Convolutional feature maps" (NoCs). We discover that aside from deep feature maps, a deep and convolutional per-region classifier is of particular importance for object detection, whereas latest superior image classification models (such as ResNets and GoogLeNets) do not directly lead to good detection accuracy without using such a per-region classifier. We show by experiments that despite the effective ResNets and Faster R-CNN systems, the design of NoCs is an essential element for the 1st-place winning entries in ImageNet and MS COCO challenges 2015.

Overview

CoChat Summary

TL;DR

The paper studies region-wise classifier networks, called Networks on Convolutional feature maps (NoCs), built on shared convolutional features for object detection. It finds that deep, convolutional per-region classifiers substantially improve detection and are essential for Faster R-CNN with ResNets to achieve strong MS COCO results.

Key findings

  • On PASCAL VOC 2007 with ZF features and VOC07 trainval, replacing SVM-on-RoI (45.8% mAP) with deeper MLP NoCs improves mAP: 2fc 49.0, 3fc 53.1, 4fc 53.6; the 4-fc NoC is 7.8% higher than the SVM.
  • Adding convolutional layers to NoCs helps with more training data: on VOC07+12, 3fc NoC reaches 56.5, 1conv3fc 58.5, 2conv3fc 58.9, and 3conv3fc 58.8; maxout scale selection raises 2conv3fc from 58.9 to 60.7 (up to 1.8% mAP). Fine-tuning gives 62.9 vs. 60.7 for a randomly initialized NoC and 47.7 for SVM on RoI (ZF, VOC07+12).
  • Deep classifiers complement deep features: VGG16 3fc is 64.6, VGG19 3fc 65.1 (+0.5), VGG16 2conv3fc 66.1 (+1.5), and VGG16 maxout 2conv3fc 68.8; with bounding box regression, NoC reaches 71.6 mAP on VOC07 and 68.8 on VOC2012. Error analysis shows stronger NoCs reduce localization error (22.6 vs. 28.1 with ZF; 20.1 vs. 24.8 with VGG16).
  • On MS COCO val with Faster R-CNN, naive ResNet-101 has 16.9 AP (stride 32) or 21.3 AP (stride 16); a 3fc MLP NoC raises these to 21.2 and 26.3, while a convolutional NoC reaches 27.2 AP, with AP@0.75 improving from 25.9 to 27.6 and AP@0.5 nearly unchanged from 48.1 to 48.4. This system is described as the foundation of the ResNet paper’s 1st-place MS COCO and ImageNet 2015 entries.

Methods

  • Ablation experiments use PASCAL VOC 2007 (20 categories, mAP on 5k test images), training on VOC07 trainval (5k) or VOC07+12 trainval (16k). Shared convolutional features from ImageNet-pretrained ZF and VGG-16/19 are frozen; ~2,000 Selective Search proposals are RoI-pooled to 6×6 (ZF) or 7×7 (VGG); NoCs are trained by backpropagation/SGD, then second-to-last fc features train a linear SVM; no bounding box regression is used in these ablations.
  • NoC families include MLPs (2–4 fc layers), ConvNets (1–3 additional 3×3 conv layers plus 3 fc layers), and maxout NoCs for scale selection from two adjacent feature-pyramid scales; fine-tuning initializes fc layers from pretrained weights and conv layers as identity mappings.
  • Faster R-CNN + ResNet experiments use MS COCO (80 categories, 80k train, 40k val), reporting COCO-style AP, AP@0.5, and AP@0.75; they compare VGG-16, GoogleNet BN-Inception, and ResNet-101 with end-to-end fine-tuning, RPN proposals, and a softmax classifier plus jointly learned bounding box regressor.

Limitations

  • Stated: the ablation experiments are mainly based on SPPnet with frozen shared feature maps, precomputed Selective Search proposals, and a post-hoc SVM; the authors say observations generally hold when these restricted conditions are relaxed or removed, as shown with Faster R-CNN. They also state VOC07 trainval is too small to train deeper models, causing overfitting, and that no bounding box regression is used in the ablation tables.
  • Stated: differences among maxout variants are marginal, and the benefit of using a ConvNet as NoC diminishes for Faster R-CNN + ResNet on PASCAL VOC (both MLP and ConvNet NoCs reach 76.4 vs. 71.9 for the 1fc baseline).
  • Evident from the text: evaluation is limited to PASCAL VOC and MS COCO with specific backbones (ZF, VGG, GoogleNet, ResNet), so broader datasets or architectures are not tested.

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

Integrity signals

Registry

Crossref · 10.1109/tpami.2016.2601099

Cites

52 of 52 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

2 Valid · 0 Not valid · 13 Not checked · 34 Own results and interpretations

22 more claims

Create a free account to see all 49 claims and to verify the 12 references still open.

See all claims — free account
  • It may state two separate claims as one. 18
  • Probably not a substantial claim. 8

Create a free account to see these claims.

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.

1 of 1

Citations

Cites

Show all 52 references

Research works printed but not linked: 1

  1. goodfellow, 2013, Maxout networks, ICML (3), 28, 1319

Cited by

Metadata

Authors
Ren, S., He, K., Girshick, R., Zhang, X., & Sun, J.
Affiliations
  • University of Science and Technology of China (CN)
  • Meta (United States) (US)
  • Xi'an Jiaotong University (CN)
Venue
IEEE Transactions on Pattern Analysis and Machine Intelligence
Published
2017-07-01
Volume
39
Issue
7
Pages
1476-1481
ISSN
  • 0162-8828 (ISSN-L)
  • 2160-9292
OpenAlex
W2117287331
Cited by
454
References
52
Citations by year
2026: 6 · 2025: 19 · 2024: 28 · 2023: 46 · 2022: 41 · 2021: 64 · 2020: 59 · 2019: 76 · 2018: 48 · 2017: 41
FWCI
42.31
Type
journal article
Language
en
Open access
Locations
  • IEEE Transactions on Pattern Analysis and Machine Intelligence (publishedVersion)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
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 › Computer Vision and Pattern Recognition › Visual Attention and Saliency Detection
Keywords
object detection, object classification, Faster R-CNN, ResNet, GoogLeNet, feature extractor, MS COCO, ImageNet

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.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-09-27, under version 3 of the claim definition (prompt extract-v3).

A second model labelled each claim on 2026-09-27 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

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

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.