Research map / Detection and segmentation / CNN features
Recurrent Scale Approximation for Object Detection in CNN
Yu Liu, Hongyang Li, Junjie Yan, Fangyin Wei, Xiaogang Wang, Xiaoou Tang
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 RSA unit to LRN for approximating smaller-scale feature maps (ablation) | LRN · AFW/FDDB/MALF average | test runtime (ms per image): 28.78 -> 14.50 | Improved |
| use LRN+RSA instead of single-anchor RPN baseline | RPN_s · FDDB | recall at 1% fpi (%): 90.61% -> 91.92% | Improved |
| use LRN+RSA instead of single-anchor RPN baseline | RPN_s · MALF | recall at 1% fpi (%): 88.81% -> 90.09% | Improved |
| use LRN+RSA instead of single-anchor RPN baseline | RPN_s · AFW | AP (%): 99.90% -> 99.96% | Improved |
| use single-anchor multi-shot design instead of multi-anchor RPN | RPN_m · FDDB | recall at 1% fpi (%): 86.89% -> 90.61% | Improved |
| use LRN+RSA instead of multi-anchor RPN | RPN_m · MALF | recall at 1% fpi (%): 84.65% -> 90.09% | Improved |
| plug RSA at deeper layer (final feature, after res3c) instead of res2b (ablation) | RSA unit · validation set | error rate: 3.44% -> almost 100% | Hurt |
| roll out RSA five times instead of once (ablation) | RSA unit at res2b · validation set | error rate: 3.44% -> 5.9% | Hurt |
| supervise landmark features with regression predictions instead of ground-truth targets (ablation) | LRN · face detection | performance | Improved |
| use RSA+RPN instead of original 18-anchor RPN | original RPN · ILSVRC DET val2 (100 proposals) | recall (%): 88.7 -> 89.1 | Improved |
| use RSA+RPN instead of single-scale RPN | single-scale RPN · ILSVRC DET val2 | speed (ms per image): 249 ms -> 124 ms | Improved |
| use RSA+RPN instead of original RPN | original RPN · ILSVRC DET val2 | speed (ms per image): 158 ms -> 124 ms | Improved |
Closest papers by meaning
- ScaleNet - Improve CNNs through Recursively Rescaling Objects (AAAI 2020 · similarity 0.88)
- Accurate Single Stage Detector Using Recurrent Rolling Convolution (CVPR 2017 · similarity 0.86)
- An Analysis of Scale Invariance in Object Detection SNIP (CVPR 2018 · similarity 0.85)
- ScaleNet: A Shallow Architecture for Scale Estimation (CVPR 2022 · similarity 0.85)
- Scale-Equalizing Pyramid Convolution for Object Detection (CVPR 2020 · similarity 0.85)
- Feature Pyramid Networks for Object Detection (CVPR 2017 · similarity 0.85)
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.
