Research map / Label-efficient and robust learning / Knowledge distillation
UniKD: Universal Knowledge Distillation for Mimicking Homogeneous or Heterogeneous Object Detectors
Shanshan Lao, Guanglu Song, Boxiao Liu, Yu Liu, Yujiu Yang
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 both content and position queries instead of one type (ablation) | UniKD · COCO RetinaNet R50 -> R18 | mAP: 34.9 / 34.5 -> 35.3 | Improved |
| combine UniKD with FitNet feature mimicking (ablation) | UniKD · COCO RetinaNet R50 -> R18 | mAP: 35.3 -> 34.9 | Hurt |
| increase queries from 200 to 300 (ablation) | UniKD · COCO RetinaNet R18 student | mAP: 35.3 -> 35.2 | No clear effect |
| use 3 cross-attention layers instead of 1 in fe (ablation) | UniKD · COCO RetinaNet R18 student | mAP: 35.3 -> 34.9 | Hurt |
| pretrain AKE for 15 instead of 5 epochs (ablation) | UniKD · COCO RetinaNet R18 student | mAP: 35.2 -> 35.3 | No clear effect |
| add 2 self-attention adaptor layers (ablation) | UniKD with Deformable DETR teacher · COCO RetinaNet R18 student | mAP: 31.9 -> 33.8 | Improved |
| add self-attention adaptor layers to FitNet (ablation) | FitNet with Deformable DETR teacher · COCO RetinaNet R18 student | mAP: 31.8 -> 31.3 | Hurt |
| replace feature mimicking with UniKD for heterogeneous teacher | RetinaNet R18 student · COCO heterogeneous pairs | mAP: FGD 31.3 -> UniKD 34.8 | Improved |
| use FitNet with Deformable DETR teacher | RetinaNet R18 student · COCO | mAP | Hurt |
| use a stronger teacher with different architecture | UniKD · COCO Table 5 | mAP | No clear effect |
Later papers that compared with this method
1 later accepted paper reports UNI-KD as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Reinforced Cross-Domain Knowledge Distillation on Time Series Data (NeurIPS 2024 · 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.
- HEAD — HEAD: HEtero-Assists Distillation for Heterogeneous Object Detectors (ECCV 2022) · better on 2 of 2 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
- HEAD: HEtero-Assists Distillation for Heterogeneous Object Detectors (ECCV 2022 · similarity 0.91)
- ScaleKD: Distilling Scale-Aware Knowledge in Small Object Detector (CVPR 2023 · similarity 0.90)
- Distilling Object Detectors with Global Knowledge (ECCV 2022 · similarity 0.90)
- Learning Lightweight Object Detectors via Multi-Teacher Progressive Distillation (ICML 2023 · similarity 0.90)
- DetKDS: Knowledge Distillation Search for Object Detectors (ICML 2024 · similarity 0.90)
- CrossKD: Cross-Head Knowledge Distillation for Object Detection (CVPR 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.
