Research map / Detection and segmentation / CNN features
Revisiting RCNN: On Awakening the Classification Power of Faster RCNN
Bowen Cheng, Yunchao Wei, Honghui Shi, Rogerio Feris, Jinjun Xiong, Thomas Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add DCR module (ResNet-50) trained with FP+FG+BG sampling (ablation) | Faster RCNN ResNet-101 · PASCAL VOC2007 test (ablation) | mAP: 79.8 -> 82.3 | Improved |
| add DCR module to Deformable Faster RCNN (ablation) | DCN · PASCAL VOC2007 test (ablation) | mAP: 81.4 -> 83.2 | Improved |
| sample hard false positives, foreground and background uniformly instead of hard FP only (ablation) | DCR · PASCAL VOC2007 test (ablation) | mAP: 81.4 -> 82.3 | Improved |
| add background samples to hard false positives (ablation) | DCR · PASCAL VOC2007 test (ablation) | mAP: 81.4 -> 80.3 | Hurt |
| replace RCNN-like fg:bg=1:3 sampling with uniform hard FP+FG+BG sampling (ablation) | DCR · PASCAL VOC2007 test (ablation) | mAP: 81.7 -> 82.3 | Improved |
| vary hard false positive threshold between 0.2 and 0.4 (ablation) | DCR · PASCAL VOC2007 test (ablation) | mAP: 82.2 -> 82.3 | No clear effect |
| vary training sample size from 8 to 64 boxes (ablation) | DCR · PASCAL VOC2007 test (ablation) | mAP: 82.0 -> 82.3 | No clear effect |
| increase DCR ROI scale from 56x56 to 224x224 (ablation) | DCR · PASCAL VOC2007 test (ablation) | mAP: 80.6 -> 82.3 | Improved |
| deepen DCR classifier from ResNet-18 to ResNet-152 (ablation) | DCR · PASCAL VOC2007 test (ablation) | mAP: 81.4 -> 82.5 | Improved |
| add DCR-50 instead of switching Faster RCNN backbone to ResNet-152 (ablation) | Faster RCNN ResNet-101 · PASCAL VOC2007 test (ablation) | mAP: 80.3 -> 82.3 | Improved |
| add DCR-50 instead of ensembling Faster RCNN ResNet-101 and ResNet-152 (ablation) | Faster RCNN ResNet-101 · PASCAL VOC2007 test (ablation) | mAP: 81.1 -> 82.3 | Improved |
| multiply RCNN classification score with Faster RCNN score (ablation) | Faster RCNN · PASCAL VOC2007 test (ablation) | mAP: 79.8 -> 81.7 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- R-FCN — R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning Rich Features at High-Speed for Single-Shot Object Detection (ICCV 2019 · similarity 0.87)
- Rethinking Classification and Localization for Object Detection (CVPR 2020 · similarity 0.87)
- Reasoning-RCNN: Unifying Adaptive Global Reasoning Into Large-Scale Object Detection (CVPR 2019 · similarity 0.87)
- Object Detection Networks on Convolutional Feature Maps (TPAMI 2016 · similarity 0.87)
- R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016 · similarity 0.87)
- Correlation Loss: Enforcing Correlation between Classification and Localization (AAAI 2023 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
