Research map / Detection and segmentation / CNN features
Bucketed Ranking-based Losses for Efficient Training of Object Detectors
Feyza Yavuz, Baris Can Cam, Adnan Harun Dogan, Kemal Oksuz, Emre Akbas, Sinan Kalkan
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 |
|---|---|---|---|
| replace RS Loss with bucketed BRS Loss | Cascade R-CNN R-50 · COCO val | iteration time (s): 1.28 -> 0.24 | Improved |
| replace RS Loss with bucketed BRS Loss | Faster R-CNN R-50 · COCO val | iteration time (s): 0.50 -> 0.17 | Improved |
| replace RS Loss with bucketed BRS Loss | Faster R-CNN R-50 · COCO val | AP: 39.4 -> 39.5 | No clear effect |
| replace AP Loss with bucketed BAP Loss | ATSS R-50 · COCO val | AP: 38.1 -> 38.5 | Improved |
| replace AP Loss with bucketed BAP Loss | PAA R-50 · COCO val | AP: 37.3 -> 37.2 | No clear effect |
| replace RS Loss with BRS Loss on instance segmentation | Mask R-CNN · COCO, Cityscapes, LVIS | iteration time (s): up to ~2.5x faster | Improved |
| spend saved time on longer training (36 epochs) with BRS vs RS at 12 epochs | Cascade R-CNN · COCO val | AP: 41.1 -> 42.3 | Improved |
| spend saved time on longer training (27 epochs) with BRS vs RS at 12 epochs | Mask R-CNN · COCO val | mask AP: 36.3 -> 37.3 | Improved |
| bucket negatives in AP Loss | AP Loss on synthetic logits · synthetic data | loss runtime: up to ~40x speed-up | Improved |
| replace Co-DETR training loss with BRS Loss | Co-DETR ResNet50 · COCO val | AP: 49.3 -> 50.1 | Improved |
Closest papers by meaning
- Rank & Sort Loss for Object Detection and Instance Segmentation (ICCV 2021 · similarity 0.92)
- DR Loss: Improving Object Detection by Distributional Ranking (CVPR 2020 · similarity 0.91)
- RankDetNet: Delving Into Ranking Constraints for Object Detection (CVPR 2021 · similarity 0.90)
- Revisiting AP Loss for Dense Object Detection: Adaptive Ranking Pair Selection (CVPR 2022 · similarity 0.90)
- AP-Loss for Accurate One-Stage Object Detection (TPAMI 2020 · similarity 0.89)
- Towards Accurate One-Stage Object Detection With AP-Loss (CVPR 2019 · similarity 0.88)
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.
