Research map / Label-efficient and robust learning / Knowledge distillation
HEAD: HEtero-Assists Distillation for Heterogeneous Object Detectors
Luting Wang, Xiaojie Li, Yue Liao, Zeren Jiang, Jianlong Wu, Fei Wang, Chen Qian, Si Liu
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 CKD loss to HEAD (AKD only) (ablation) | HEAD, R18 RetinaNet, teacher R50 Faster R-CNN · COCO val, R18 RetinaNet student, teacher R50 Faster R-CNN | mAP: 35.5 -> 36.2 | Improved |
| add CKD loss to HEAD (AKD only) (ablation) | HEAD, MNV2 RetinaNet, teacher R50 Faster R-CNN · COCO val, MNV2 RetinaNet student, teacher R50 Faster R-CNN | mAP: 32.5 -> 32.8 | Improved |
| add CKD loss to HEAD (AKD only) (ablation) | HEAD, R18 RetinaNet, teacher R50 RepPoints · COCO val, R18 RetinaNet student, teacher R50 RepPoints | mAP: 34.2 -> 34.3 | No clear effect |
| add assistant-based KD from R50 Faster R-CNN | R18 RetinaNet · COCO val, R18 RetinaNet student, teacher R50 Faster R-CNN | mAP: 31.7 -> 35.5 | Improved |
| use HEAD instead of homo-KD FitNet with RepPoints teacher | R18 RetinaNet · COCO val, R18 RetinaNet student, teacher R50 RepPoints | mAP: 31.5 -> 34.2 | Improved |
| apply homo-KD DeFeat with RepPoints teacher | R18 RetinaNet · COCO val, R18 RetinaNet student, teacher R50 RepPoints | mAP: 31.7 -> 30.9 | Hurt |
| use HEAD instead of G-DetKD | R18 RetinaNet, teacher R50 Faster R-CNN · COCO val, R18 RetinaNet student, teacher R50 Faster R-CNN | mAP: 35.4 -> 36.2 | Improved |
| use HEAD instead of G-DetKD | R18 FCOS, teacher R50 Faster R-CNN · COCO val, R18 FCOS student, teacher R50 Faster R-CNN | mAP: 34.1 -> 36.0 | Improved |
| use HEAD instead of G-DetKD | MNV2 RetinaNet, teacher R50 Faster R-CNN · COCO val, MNV2 RetinaNet student, teacher R50 Faster R-CNN | mAP: 32.6 -> 32.8 | No clear effect |
| add FCOS assistant to RetinaNet (teacher-free) (ablation) | TF-HEAD, R18 RetinaNet · COCO val, TF-HEAD ablation, R18 RetinaNet | mAP: 31.7 -> 32.9 | Improved |
| add CKD loss to FCOS assistant (ablation) | TF-HEAD, R18 RetinaNet · COCO val, TF-HEAD ablation, R18 RetinaNet | mAP: 32.9 -> 33.0 | No clear effect |
| add R-CNN assistant to RetinaNet (teacher-free) (ablation) | TF-HEAD, R18 RetinaNet · COCO val, TF-HEAD ablation, R18 RetinaNet | mAP: 31.7 -> 33.4 | Improved |
Later papers that compared with this method
1 later accepted paper reports HEAD as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- UniKD: Universal Knowledge Distillation for Mimicking Homogeneous or Heterogeneous Object Detectors (ICCV 2023 · better on 2 of 2 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- G-DetKD — G-DetKD: Towards General Distillation Framework for Object Detectors via Contrastive and Semantic-Guided Feature Imitation (ICCV 2021) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PKD: General Distillation Framework for Object Detectors via Pearson Correlation Coefficient (NeurIPS 2022 · similarity 0.94)
- UniKD: Universal Knowledge Distillation for Mimicking Homogeneous or Heterogeneous Object Detectors (ICCV 2023 · similarity 0.91)
- Distilling Object Detectors via Monte Carlo Dropout (TPAMI 2026 · similarity 0.91)
- DCSF-KD: Dynamic Channel-wise Spatial Feature Knowledge Distillation for Object Detection (AAAI 2025 · similarity 0.91)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.90)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · 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.
