Research map / Detection and segmentation / CNN features
Loss Function Discovery for Object Detection via Convergence-Simulation Driven Search
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 default CE + L1 losses with CSE-Autoloss-A | Faster R-CNN R50 · COCO val | AP (%): 37.4 -> 38.5 | Helped |
| Transfer CSE-Autoloss-A to different two-stage detectors | Faster R-CNN R101, Cascade R-CNN R50, and Mask R-CNN R50 · COCO val | AP (%): 39.4 -> 40.2; 40.3 -> 40.5; 38.2 -> 39.1 | Helped |
| Replace FL + GIoU with CSE-Autoloss-B | FCOS R50 · COCO val | AP (%): 38.8 -> 39.6 | Helped |
| Transfer CSE-Autoloss-B to ATSS | ATSS R50 · COCO val | AP (%): 40.0 -> 40.5 | Helped |
| Transfer CSE-Autoloss-A to other datasets | Faster R-CNN · PASCAL VOC 2007 test and BDD val | mAP (%) / AP (%): 79.5 -> 80.4; 36.5 -> 37.3 | Helped |
| Add IoU weighting to CE to form CEI (ablation) | Faster R-CNN R50 · COCO val | AP (%): 37.6 -> 37.7 | Helped |
| Add IoU weighting to focal loss to form FLI (ablation) | FCOS R50 · COCO val | AP (%): 38.8 -> 39.0 | Helped |
| Use the searched regression loss with default classification loss (ablation) | Faster R-CNN R50 · COCO val | AP (%): 37.4 -> 37.9 | Helped |
| Use the searched classification loss with default GIoU regression (ablation) | Faster R-CNN R50 · COCO val | AP (%): 37.6 -> 38.2 | Helped |
| Use the searched regression loss with focal classification (ablation) | FCOS R50 · COCO val | AP (%): 38.8 -> 39.1 | Helped |
| Use the searched classification loss with default GIoU regression (ablation) | FCOS R50 · COCO val | AP (%): 38.8 -> 39.4 | Helped |
| Add convergence property verification to candidate filtering (ablation) | CSE-Autoloss search · Faster R-CNN classification loss search | Evaluated loss candidates: $5\times 10^3$ -> $7\times 10^2$ | Helped |
Closest papers by meaning
- Loss Function Search for Face Recognition (ICML 2020 · similarity 0.90)
- AutoLoss-Zero: Searching Loss Functions From Scratch for Generic Tasks (CVPR 2022 · similarity 0.89)
- $\alpha$-IoU: A Family of Power Intersection over Union Losses for Bounding Box Regression (NeurIPS 2021 · similarity 0.87)
- Searching Parameterized AP Loss for Object Detection (NeurIPS 2021 · similarity 0.86)
- Supervision Interpolation via LossMix: Generalizing Mixup for Object Detection and Beyond (AAAI 2024 · similarity 0.85)
- YOLOv10: Real-Time End-to-End Object Detection (NeurIPS 2024 · 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.