Research map / Detection and segmentation / CNN features
A Ranking-based, Balanced Loss Function Unifying Classification and Localisation in Object Detection
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 AP Loss and Smooth L1 with aLRP Loss | RetinaNet, ResNet-50 · COCO minival | AP: 35.5 to 40.2 | Helped |
| add ranking-based localisation to classification ranking (ablation) | aLRP Loss on RetinaNet · COCO minival | AP: 35.5 to 36.9 | Helped |
| add self-balancing (ablation) | aLRP Loss with IoU localisation · COCO minival | AP: 36.9 to 38.7 | Helped |
| replace IoU with GIoU (ablation) | aLRP Loss with self-balancing · COCO minival | AP: 38.7 to 38.9 | Helped |
| use ATSS assignment (ablation) | aLRP Loss with GIoU and self-balancing · COCO minival | AP: 38.9 to 40.2 | Helped |
| replace self-balancing with constant weighting, w_r=10 (ablation) | aLRP Loss with IoU localisation · COCO minival | AP: 38.7 to 38.6 | Hurt |
| replace self-balancing with constant weighting, w_r=15 (ablation) | aLRP Loss with GIoU localisation · COCO minival | AP: 38.9 to 38.8 | Hurt |
| vary self-balancing initial weight from 1 to 500 (ablation) | aLRP Loss · COCO minival | AP: 38.8; 38.9; 38.7; 38.5 | No clear effect |
| align localisation ranking with classification ranking rather than oppose it | aLRP Loss · COCO minival ranking bounds | AP: 28.6 to 48.1 | Helped |
| train with aLRP Loss instead of Focal Loss+SL1 | FoveaBox, ResNet-50 · COCO minival | AP: 38.3 to 39.7 | Helped |
| train with aLRP Loss instead of AP Loss+SL1 | FoveaBox, ResNet-50 · COCO minival | AP: 36.5 to 39.7 | Helped |
| replace Cross Entropy+GIoU with aLRP Loss | Faster R-CNN, ResNet-50 · COCO minival | AP: 38.2 to 40.7 | Helped |
Closest papers by meaning
- Revisiting AP Loss for Dense Object Detection: Adaptive Ranking Pair Selection (CVPR 2022 · similarity 0.88)
- Searching Parameterized AP Loss for Object Detection (NeurIPS 2021 · similarity 0.87)
- Towards Accurate One-Stage Object Detection With AP-Loss (CVPR 2019 · similarity 0.86)
- Correlation Loss: Enforcing Correlation between Classification and Localization (AAAI 2023 · similarity 0.85)
- DCA: Dividing and Conquering Amnesia in Incremental Object Detection (AAAI 2025 · similarity 0.82)
- How to Evaluate the Generalization of Detection? A Benchmark for Comprehensive Open-Vocabulary Detection (AAAI 2024 · similarity 0.82)
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.