Research map / Detection and segmentation / CNN features
Revisiting AP Loss for Dense Object Detection: Adaptive Ranking Pair Selection
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 |
|---|---|---|---|
| Add adaptive positive-positive ranking pairs using localization quality (APE) (ablation) | AP Loss on RetinaNet · MS-COCO val-2017 | AP: 37.3 -> 38.3 | Helped |
| Select ranking pairs with ATSS instead of an IoU threshold (ablation) | Pairwise Error on RetinaNet · MS-COCO val-2017 | AP: 37.3 to 39.2 | Helped |
| Use normalized ranking and localization scores together for PAA clustering (ablation) | APE with PAA · MS-COCO val-2017 | AP: 38.9 and 33.9; 2.2 AP performance gain | Helped |
| Add 0-1 normalization to ranking and localization scores (ablation) | APE with PAA clustering · MS-COCO val-2017 | AP: 40.5 to 41.1; 0.6 AP performance gain | Helped |
| Replace normalized ranking and localization scores with Focal Loss and GIoU Loss inputs (ablation) | APE with PAA clustering · MS-COCO val-2017 | AP: 41.1 to 39.4 | Hurt |
| Use ranking scores alone for clustering instead of pairing ranking and localization scores (ablation) | APE with PAA clustering · MS-COCO val-2017 | AP: 38.9; 2.2 AP performance gain from associating both scores | Hurt |
| Use localization scores alone for clustering instead of pairing ranking and localization scores (ablation) | APE with PAA clustering · MS-COCO val-2017 | AP: 33.9; 2.2 AP performance gain from associating both scores | Hurt |
| Replace the step distance H with sigmoid distance S (ablation) | AP Loss · MS-COCO val-2017 | AP: 37.4 -> 37.3 | No clear effect |
| Replace Error-Driven Update with cross-entropy loss (ablation) | AP Loss · MS-COCO val-2017 | AP: 37.3 | No clear effect |
| Replace adaptive AP balance constant with the valid-negative count N_neg (ablation) | Pairwise Error · MS-COCO val-2017 | AP: 37.3 v.s. 37.3 | No clear effect |
| Use FCOS(800) instead of RetinaNet(512) (ablation) | APE · MS-COCO val-2017 | AP: training on FCOS(800) can improve our method by 0.4 AP | Helped |
| Train the PAA* APE model for 96 instead of 48 epochs (ablation) | APE, PAA*(512) · MS-COCO val-2017 | AP: 41.1 -> 42.9 | Helped |
Closest papers by meaning
- Searching Parameterized AP Loss for Object Detection (NeurIPS 2021 · similarity 0.89)
- DR Loss: Improving Object Detection by Distributional Ranking (CVPR 2020 · similarity 0.89)
- Towards Accurate One-Stage Object Detection With AP-Loss (CVPR 2019 · similarity 0.88)
- How to Evaluate the Generalization of Detection? A Benchmark for Comprehensive Open-Vocabulary Detection (AAAI 2024 · similarity 0.83)
- Plug and Play Active Learning for Object Detection (CVPR 2024 · similarity 0.83)
- RankED: Addressing Imbalance and Uncertainty in Edge Detection Using Ranking-based Losses (CVPR 2024 · similarity 0.83)
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.