Research map / Label-efficient and robust learning / Knowledge distillation
G-DetKD: Towards General Distillation Framework for Object Detectors via Contrastive and Semantic-Guided Feature Imitation
Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, Tong Zhang
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 |
|---|---|---|---|
| use SGFI instead of RoI feature imitation (ablation) | FasterRCNN-R18 student · COCO val | AP: 36.4 -> 36.7 | Improved |
| use SGFI instead of GT-mask imitation (ablation) | RetinaNet-R18 student · COCO val | AP: 34.3 -> 36.0 | Improved |
| add contrastive KD on top of SGFI (ablation) | FasterRCNN-R18 student · COCO val | AP: 36.7 -> 37.6 | Improved |
| add prediction KD and head transfer to SGFI+CKD (ablation) | FasterRCNN-R18 student · COCO val | AP: 37.6 -> 37.9 | Improved |
| add prediction KD to SGFI (ablation) | RetinaNet-R18 student, FasterRCNN-R50 teacher · COCO val | AP: 36.1 -> 36.1 | No clear effect |
| use class-aware regression KD instead of class-agnostic (ablation) | FasterRCNN-R18 student · COCO val | AP: 34.7 -> 35.7 | Improved |
| increase contrastive memory queue from 8x1024 to 80x1024 (ablation) | CKD FasterRCNN-R18 · COCO val | AP: 36.8 -> 37.1 | Improved |
| raise IoU threshold for negative assignment from 0.1 to 0.5 (ablation) | CKD FasterRCNN-R18 · COCO val | AP: 36.8 -> 37.1 | Improved |
| use linear instead of nonlinear projection head (ablation) | CKD FasterRCNN-R18 · COCO val | AP: 36.7 -> 37.1 | Improved |
| form contrastive pairs from classification branch instead of localization branch (ablation) | CKD RetinaNet-R18 · COCO val | AP: 34.3 -> 35.8 | Improved |
| distill from a weaker R18 teacher into R50 student | G-DetKD R50 student · COCO val | AP: 37.4 -> 37.7 | Improved |
| apply G-DetKD to R101 student with R50 teacher | G-DetKD · COCO val | AP: 39.4 -> 40.4 | Improved |
Later papers that compared with this method
1 later accepted paper reports G-DetKD as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- HEAD: HEtero-Assists Distillation for Heterogeneous Object Detectors (ECCV 2022 · 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
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.85)
- Knowledge Distillation for Object Detection via Rank Mimicking and Prediction-Guided Feature Imitation (AAAI 2022 · similarity 0.84)
- Distilling Object Detectors With Fine-Grained Feature Imitation (CVPR 2019 · similarity 0.84)
- IMC-Det: Intra–Inter Modality Contrastive Learning for Video Object Detection (IJCV 2024 · similarity 0.83)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 2024 · similarity 0.82)
- GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation (AAAI 2025 · similarity 0.82)
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.
