Research map / Detection and segmentation / CNN features
Sparse R-CNN: End-to-End Object Detection With Learnable Proposals
Check 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 the dense RPN with learnable sparse proposal boxes (ablation) | Faster R-CNN · COCO 2017 val | AP: 40.2 -> 18.5 | Hurt |
| Add iterative box refinement (ablation) | Sparse R-CNN with learnable proposals · COCO 2017 val | AP: +13.7 AP | Helped |
| Concatenate previous-stage object features into the current stage (ablation) | Sparse R-CNN iterative architecture · COCO 2017 val | AP: +11.7 AP | Helped |
| Add self-attention to object features (ablation) | Sparse R-CNN iterative model · COCO 2017 val | AP: 32.2 -> 37.2 (+5.0) | Helped |
| Add dynamic instance interaction after self-attention (ablation) | Sparse R-CNN · COCO 2017 val | AP: 37.2 -> 42.3 (+5.1) | Helped |
| Replace multi-head attention with the dynamic head (ablation) | Sparse R-CNN recognition head · COCO 2017 val | AP: +6.6 AP | Helped |
| Change proposal-box initialization among center, whole-image, grid, and random (ablation) | Sparse R-CNN · COCO 2017 val | AP: 41.0 to 42.3 | No clear effect |
| Increase the number of proposals from 100 to 500 (ablation) | Sparse R-CNN · COCO 2017 val | AP: 42.3 -> 44.6 | Helped |
| Increase proposals from 100 to 500 (ablation) | Sparse R-CNN · COCO 2017 val | training time: 19h -> 60h | Hurt |
| Increase iterative stage count from 1 to 6 (ablation) | Sparse R-CNN · COCO 2017 val | AP: 21.7 -> 42.3 | Helped |
| Increase iterative stage count from 6 to 12 (ablation) | Sparse R-CNN · COCO 2017 val | AP: 42.3 -> 41.6 | Hurt |
| Remove positional encoding from the proposal features (ablation) | Sparse R-CNN · COCO 2017 val | AP: 41.9 -> 42.3 | No clear effect |
Closest papers by meaning
- Dynamic Sparse R-CNN (CVPR 2022 · similarity 0.93)
- R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016 · similarity 0.91)
- Cascade R-CNN: Delving Into High Quality Object Detection (CVPR 2018 · similarity 0.91)
- Sparse DETR: Efficient End-to-End Object Detection with Learnable Sparsity (ICLR 2022 · similarity 0.90)
- Spatial Pruned Sparse Convolution for Efficient 3D Object Detection (NeurIPS 2022 · similarity 0.90)
- SparseFormer: Sparse Visual Recognition via Limited Latent Tokens (ICLR 2024 · 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.