Research map / Detection and segmentation / Weakly supervised detection and segmentation
Spatial Self-Distillation for Object Detection with Inaccurate Bounding Boxes
Di Wu, Pengfei Chen, Xuehui Yu, Guorong Li, Zhenjun Han, Jianbin Jiao
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 Spatial Position Self-Distillation (SPSD) bag construction (ablation) | SSD-Det · MS-COCO val, 40% noise | AP: 24.6 -> 27.2 | Improved |
| add Spatial Identity Self-Distillation (SISD) for proposal selection (ablation) | SSD-Det · MS-COCO val, 40% and 20% noise | AP: 27.2 -> 27.6 (40%), 33.0 -> 33.6 (20%) | Improved |
| re-train Faster R-CNN on SSD-Det pseudo labels after end-to-end training (ablation) | SSD-Det · MS-COCO val, 40% noise | AP: 27.6 -> 29.3 | Improved |
| end-to-end joint training before re-train instead of training only the refiner (ablation) | SSD-Det + Re-Train · MS-COCO val, 40% noise | AP: 29.0 -> 29.3 | Improved |
| weight SPSD proposals with stage I score (ablation) | SPSD · MS-COCO val, 40% noise | AP: 26.0 -> 26.3 | No clear effect |
| use SPSD in both stages instead of only stage II (ablation) | SPSD · MS-COCO val, 40% noise | AP: 26.3 -> 27.2 | Improved |
| increase SPSD modules from two to three (ablation) | SSD-Det · MS-COCO val, 40% noise | AP: 27.2 -> 27.0 | Hurt |
| subtract or concatenate object and proposal features in SISD ORE (ablation) | SISD · MS-COCO val, 40% noise | AP: 27.2 (subtract), 27.2 (concat) versus 27.6 (add) | No clear effect |
| share the two fc layers in SISD (ablation) | SISD · MS-COCO val, 40% noise | AP: 27.6 -> 23.0 | Hurt |
| re-train advanced detectors on SSD-Det refined boxes | Deformable DETR · MS-COCO, 40% noise | AP: 5.0 -> 35.2 | Improved |
| use spatial self-distillation instead of MIL refinement | OA-MIL · MS-COCO val, 40% noise | drift rate (%): 15.1 -> 1.5 | Improved |
| SSD-Det at low noise on VOC compared with OA-MIL | OA-MIL · VOC 2007, 10% noise | AP50: 77.4 -> 77.1 | No clear effect |
Closest papers by meaning
- Multi-Task Self-Supervised Object Detection via Recycling of Bounding Box Annotations (CVPR 2019 · similarity 0.86)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.86)
- Instance Localization for Self-Supervised Detection Pretraining (CVPR 2021 · similarity 0.86)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.85)
- DIOD: Self-Distillation Meets Object Discovery (CVPR 2024 · similarity 0.85)
- Self-Supervised Learning of Intertwined Content and Positional Features for Object Detection (ICML 2025 · similarity 0.84)
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.
