Research map / Detection and segmentation / Open-set and few-shot detection
Boosting Long-tailed Object Detection via Step-wise Learning on Smooth-tail Data
Na Dong, Yongqiang Zhang, Mingli Ding, Gim Hee Lee
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 head-class fine-tuning step (ablation) | step-wise learning · LVIS v0.5 (ResNet-50) | AP: 27.0 -> 29.7 | Improved |
| add knowledge transfer on tail-dominant data (ablation) | step-wise learning · LVIS v0.5 (ResNet-50) | AP: 27.0 -> 29.4 | Improved |
| combine fine-tuning and knowledge transfer (ablation) | step-wise learning · LVIS v0.5 (ResNet-50) | AP: 27.0 -> 30.3 | Improved |
| combine fine-tuning and knowledge transfer (ablation) | step-wise learning · LVIS v0.5 (ResNet-50) | AP (frequent): 31.6 -> 30.9 | Hurt |
| share object query features from head expert to unified model (ablation) | knowledge transfer · LVIS v0.5 (ResNet-50) | AP: 24.4 -> 30.3 | Improved |
| add feature-level distillation loss with head-class mask (ablation) | knowledge transfer · LVIS v0.5 (ResNet-50) | AP: 29.7 -> 30.3 | Improved |
| add classification distillation loss (ablation) | knowledge transfer · LVIS v0.5 (ResNet-50) | AP: 29.4 -> 30.3 | Improved |
| use two-step division at M=30 instead of three- or four-step divisions (ablation) | step-wise learning · LVIS v0.5 (ResNet-50) | AP: 29.3 / 29.8 -> 30.3 | Improved |
| vary head/tail split threshold M among two-step divisions (10, 30, 50, 100) (ablation) | step-wise learning · LVIS v0.5 (ResNet-50) | AP: 30.1, 30.3, 30.2, 30.1 | No clear effect |
| add step-wise repeat factor sampling on replay data (ablation) | step-wise learning · LVIS v0.5 (ResNet-50) | AP (rare): 19.0 -> 24.9 | Improved |
| increase tail exemplars N_ex in head-class dominant data to large values (ablation) | step-wise learning · LVIS v0.5 (ResNet-50) | AP: 30.2 (10), 30.3 (30), 30.2 (50), 30.2 (100) | No clear effect |
Closest papers by meaning
- Adaptive Hierarchical Representation Learning for Long-Tailed Object Detection (CVPR 2022 · similarity 0.90)
- Rectify the Regression Bias in Long-Tailed Object Detection (ECCV 2024 · similarity 0.90)
- SimLTD: Simple Supervised and Semi-Supervised Long-Tailed Object Detection (CVPR 2025 · similarity 0.89)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 2024 · similarity 0.88)
- Pursuing Better Decision Boundaries for Long-Tailed Object Detection via Category Information Amount (ICLR 2025 · similarity 0.88)
- Decoupled Optimisation for Long-Tailed Visual Recognition (AAAI 2024 · similarity 0.87)
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.
