Research map / Detection and segmentation / CNN features
R-FCN: Object Detection via Region-based Fully Convolutional Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace RoI pooling after conv5 with position-sensitive score maps (k=7x7) in R-FCN (ablation) | ResNet-101 R-FCN · PASCAL VOC 2007 test | mAP (%): 76.6 vs 68.9 for naive Faster R-CNN at 7x7 RoI output | Helped |
| Remove position sensitivity (k=1, global pooling per RoI) (ablation) | R-FCN · PASCAL VOC 2007 test | mAP (%): does not converge (fail) vs 76.6 with k=7x7 | Hurt |
| Increase position-sensitive grid from 3x3 to 7x7 (ablation) | R-FCN · PASCAL VOC 2007 test | mAP (%): 75.5 to 76.6 | Helped |
| Use à trous trick (stride 16 in conv5) instead of stride 32 (ablation) | R-FCN ResNet-101 · PASCAL VOC 2007 test | mAP (%): +2.6 mAP (74.0 to 76.6) | Helped |
| Use 21-class class-specific RPN instead of class-agnostic RPN in naive setting | Fully convolutional detector with ResNet-101 · PASCAL VOC 2007 test | mAP (%): 67.6 vs 76.4 for standard Faster R-CNN | Hurt |
| Replace Faster R-CNN per-RoI 10-layer subnetwork with R-FCN negligible per-RoI computation | ResNet-101 detector · PASCAL VOC 2007 test | mAP (%) and test time: 79.5 vs 79.3 with OHEM; 76.6 vs 76.4 without; test 0.17 vs 0.42 s/img | No clear effect |
| Use multi-scale training (scales 400-800) | R-FCN ResNet-101 · PASCAL VOC 2007 test | mAP (%): 79.5 to 80.5 | Helped |
| Pretrain on MS COCO before fine-tuning on VOC | R-FCN multi-scale ResNet-101 · PASCAL VOC 2007 test | mAP (%): 80.5 to 83.6 | Helped |
| Use deeper backbone ResNet-101 instead of ResNet-50 | R-FCN · PASCAL VOC 2007 test | mAP (%): 77.0 (ResNet-50) to 79.5 (ResNet-101) | Helped |
| Increase backbone depth from ResNet-101 to ResNet-152 | R-FCN · PASCAL VOC 2007 test | mAP (%): 79.5 to 79.6 | No clear effect |
| Use OHEM with 300 RoIs during training | R-FCN · PASCAL VOC 2007 test | mAP (%): 76.6 to 79.5 (with OHEM 300 RoIs) | Helped |
| Use OHEM with 2000 RoIs instead of 300 | R-FCN · PASCAL VOC 2007 test | mAP (%): 79.3 vs 79.5 with 300 RoIs | No clear effect |
Closest papers by meaning
- R-FCN++: Towards Accurate Region-Based Fully Convolutional Networks for Object Detection (AAAI 2018 · similarity 0.95)
- Sparse R-CNN: End-to-End Object Detection With Learnable Proposals (CVPR 2021 · similarity 0.91)
- Cascade R-CNN: Delving Into High Quality Object Detection (CVPR 2018 · similarity 0.91)
- Object Detection Based on Region Decomposition and Assembly (AAAI 2019 · similarity 0.91)
- SCNN: A General Distribution Based Statistical Convolutional Neural Network with Application to Video Object Detection (AAAI 2019 · similarity 0.90)
- Spatially Invariant Unsupervised Object Detection with Convolutional Neural Networks (AAAI 2019 · similarity 0.90)
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.