Research map / Detection and segmentation / Open-set and few-shot detection
Large-Scale Object Detection in the Wild From Imbalanced Multi-Labels
Junran Peng, Xingyuan Bu, Ming Sun, Zhaoxiang Zhang, Tieniu Tan, Junjie Yan
Check 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 |
|---|---|---|---|
| Use concurrent softmax instead of softmax loss during training (ablation) | Open Images detection with class-aware sampling · Open Images mini-val | mAP: 55.45 -> 56.58 | Improved |
| Use concurrent softmax instead of softmax loss during training (ablation) | Open Images detection with non-balance sampling · Open Images mini-val | mAP: 38.16 -> 39.14 | Improved |
| Apply concurrent softmax at inference only (ablation) | Softmax-trained detector · Open Images mini-val | mAP: 55.45 -> 55.77 | Improved |
| Use BCE or focal loss rather than softmax-based losses (ablation) | Open Images multi-label detection · Open Images mini-val | mAP: 50.18; 54.29 | Hurt |
| Use soft-balance sampling with λ=0.7 rather than class-aware sampling (ablation) | Open Images detection · Open Images mini-val | mAP: 55.45 -> 57.04 | Improved |
| Set soft-balance λ to 0.3 (ablation) | Open Images detection · Open Images mini-val | mAP: 50.69 | Hurt |
| Set soft-balance λ to 1.5 (ablation) | Open Images detection · Open Images mini-val | mAP: 52.41 | Hurt |
| Extend class-aware training from 7 to 14 epochs with ImageNet pretraining (ablation) | ResNeXt-152 detector · Open Images full-val | mAP: 64.68 -> 62.85 | Hurt |
| Add class-aware finetuning after non-balance training (ablation) | ResNeXt-152 detector initialized from ImageNet · Open Images full-val | mAP: 59.85 -> 65.60 | Improved |
| Add class-aware finetuning after non-balance training from scratch (ablation) | ResNeXt-152 detector · Open Images full-val | mAP: 60.70 -> 65.92 | Improved |
| Use soft-balance rather than class-aware finetuning after non-balance training from scratch (ablation) | ResNeXt-152 detector · Open Images full-val | mAP: 65.92 -> 67.09 | Improved |
| Apply concurrent softmax in training and testing (ablation) | Soft-balance hybrid ResNeXt-152 detector · Open Images full-val | mAP: 67.09 -> 68.23 | Improved |
Closest papers by meaning
- Large-Scale Object Detection in the Wild With Imbalanced Data Distribution, and Multi-Labels (TPAMI 2024 · similarity 0.98)
- "A Simple Approach and Benchmark for 21,000-Category Object Detection" (ECCV 2022 · similarity 0.88)
- Overcoming Classifier Imbalance for Long-Tail Object Detection With Balanced Group Softmax (CVPR 2020 · similarity 0.88)
- Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty (NeurIPS 2026 · similarity 0.88)
- Diverse Learner: Exploring Diverse Supervision for Semi-Supervised Object Detection (ECCV 2022 · similarity 0.88)
- Robust Visual Recognition with Class-Imbalanced Open-World Noisy Data (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.
