Research map / Architectures and efficiency / Neural architecture search
Joint-DetNAS: Upgrade Your Detector With NAS, Pruning and Dynamic Distillation
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 |
|---|---|---|---|
| Jointly optimize architecture search, pruning, and distillation instead of applying them sequentially | R101-FPN · MS COCO | AP: 42.1$^{+3.1}$ -> 43.9; 41.8$^{+1.9}$ -> 43.9 | Helped |
| Upgrade R101-FPN with Joint-DetNAS | R101-FPN · MS COCO test-dev | AP: 41.4 -> 43.9 | Helped |
| Upgrade R101-FPN with Joint-DetNAS | R101-FPN · MS COCO test-dev | FPS: 15.9 -> 23.3 | Helped |
| Train sampled subnet with the elastic teacher pool rather than standard training (ablation) | R50-FPN · MS COCO | AP: 38.4/39.5 -> 41.8$^{+3.4}$/42.4$^{+2.9}$ | Helped |
| Prune 10% channels (ablation) | R50-FPN · MS COCO | AP: 37.1 -> 37.3 | Helped |
| Prune 20% channels (ablation) | R50-FPN · MS COCO | AP: 37.1 -> 37.0 | Hurt |
| Prune 30% channels (ablation) | R101-FPN · MS COCO | AP: 39.0 -> 38.3 | Hurt |
| Use the proposed detection KD method | R18-FPN student with R50-FPN teacher · MS COCO | AP: 36.0 -> 38.1 | Helped |
| Use proposal features rather than whole features, anchor masks, or Gaussian masks (ablation) | R18-FPN feature distillation · MS COCO | AP: 35.2$^{+1.2}$; 35.6$^{+1.6}$; 35.4$^{+1.4}$ -> 36.7$^{+2.7}$ | Helped |
| Dynamically sample a teacher instead of using the fixed largest super-net (ablation) | R18-FPN student · MS COCO | AP | Helped |
| Add class-aware localization distillation instead of directly distilling localization outputs (ablation) | R18-FPN student · MS COCO | AP: 34.8 -> 35.7 | Helped |
| Combine feature distillation, RCNN classification distillation, and class-aware localization distillation (ablation) | R18-FPN student · MS COCO | AP: 34.0 -> 37.9$^{+3.9}$ | Helped |
Closest papers by meaning
- DetNAS: Backbone Search for Object Detection (NeurIPS 2019 · similarity 0.92)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.91)
- Hybrid Proposal Refiner: Revisiting DETR Series from the Faster R-CNN Perspective (CVPR 2024 · similarity 0.87)
- Order Matters: Competition-Guided Query Ordering for RNN-Based Object Detection (NeurIPS 2026 · similarity 0.87)
- MI-DETR: An Object Detection Model with Multi-time Inquiries Mechanism (CVPR 2025 · similarity 0.87)
- MS-DETR: Efficient DETR Training with Mixed Supervision (CVPR 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.