Research map / Label-efficient and robust learning / Knowledge distillation
Deep Structured Instance Graph for Distilling Object Detectors
Yixin Chen, Pengguang Chen, Shu Liu, Liwei Wang, Jiaya Jia
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 edge (pairwise relation) distillation to the student (ablation) | Faster R-CNN R18 student · COCO val2017 Faster R-CNN R18-R50 ablation (1x) | Box AP: 33.06 -> 33.95 | Improved |
| add foreground node distillation on top of edges (ablation) | Faster R-CNN R18 student · COCO val2017 Faster R-CNN R18-R50 ablation (1x) | Box AP: 33.95 -> 36.64 | Improved |
| add background nodes with adaptive loss weight (ablation) | Faster R-CNN R18 student · COCO val2017 Faster R-CNN R18-R50 ablation (1x) | Box AP: 36.64 -> 37.17 | Improved |
| distill the structured instance graph (1x), R18 student from R50 teacher | Faster R-CNN · COCO val2017 Faster R-CNN (1x) | Box AP: 33.06 -> 37.25 | Improved |
| distill the structured instance graph (1x), R50 student from R101 teacher | Faster R-CNN · COCO val2017 Faster R-CNN (1x) | Box AP: 38.03 -> 40.57 | Improved |
| distill the structured instance graph (1x), R101 student from R152 teacher | Faster R-CNN · COCO val2017 Faster R-CNN (1x) | Box AP: 40.27 -> 41.65 | Improved |
| distill the graph into a heterogeneous MobileNetV2 student from R50 teacher | Faster R-CNN · COCO val2017 Faster R-CNN (1x) | Box AP: 29.47 -> 34.44 | Improved |
| distill the graph into RetinaNet R18 student | RetinaNet · COCO val2017 RetinaNet (1x) | Box AP: 31.60 -> 34.72 | Improved |
| train distilled R18 student for 3x instead of 1x schedule | Faster R-CNN R18 student · COCO val2017 Faster R-CNN | Box AP: 37.25 -> 38.68 | Improved |
| graph distillation vs pixel-wise pairwise distillation (R18-R50) | Mask R-CNN R18 student · COCO val2017 Mask R-CNN R18-R50 (1x) | Mask AP: 30.43 -> 33.90 | Improved |
| graph distillation vs AttentionGuided at 2x (R50-R101) | Faster R-CNN R50 student · COCO val2017 Faster R-CNN R50-R101 (Table 4 comparison, 2x) | Box AP: 41.50 -> 41.55 | No clear effect |
| graph distillation vs AttentionGuided at 2x (R18-R50) | Faster R-CNN R18 student · COCO val2017 Faster R-CNN R18-R50 (Table 4 comparison, 2x) | Box AP: 37.00 -> 38.09 | Improved |
Closest papers by meaning
- Learning Lightweight Object Detectors via Multi-Teacher Progressive Distillation (ICML 2023 · similarity 0.91)
- Distilling Object Detectors with Global Knowledge (ECCV 2022 · similarity 0.91)
- Structured Knowledge Distillation for Accurate and Efficient Object Detection (TPAMI 2023 · similarity 0.90)
- Instance-Conditional Knowledge Distillation for Object Detection (NeurIPS 2021 · similarity 0.89)
- Structure Inference Net: Object Detection Using Scene-Level Context and Instance-Level Relationships (CVPR 2018 · similarity 0.89)
- Instance-Aware Semantic Segmentation via Multi-Task Network Cascades (CVPR 2016 · similarity 0.89)
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.
