Research map / Continual learning / Class-incremental
Alleviating Catastrophic Forgetting of Incremental Object Detection via Within-Class and Between-Class Knowledge Distillation
Mengxue Kang, Jinpeng Zhang, Jinming Zhang, Xiashuang Wang, Yang Chen, Zhe Ma, Xuhui Huang
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 Interactive Feature Distillation (IFD) to the LwF baseline (ablation) | Deformable DETR with LwF · COCO last 40 + first 40 | final mAP: 38.89 -> 39.68 | Improved |
| add Distance Matrix Distillation (DMD) to the LwF baseline (ablation) | Deformable DETR with LwF · COCO last 40 + first 40 | final mAP: 38.89 -> 39.58 | Improved |
| combine DMD and IFD (ablation) | Deformable DETR with LwF · COCO last 40 + first 40 | final mAP: 38.89 -> 39.80 | Improved |
| distill the whole feature map instead of selected features (ablation) | IFD · COCO last 40 + first 40 | final mAP: 38.89 -> 38.93 | No clear effect |
| restrict feature distillation to foreground objects (ablation) | IFD · COCO last 40 + first 40 | final mAP: 38.89 -> 39.36 | Improved |
| add high-level semantic interaction to foreground feature distillation (ablation) | IFD · COCO last 40 + first 40 | final mAP: 39.36 -> 39.68 | Improved |
| use Manhattan distance instead of Euclidean in DMD (ablation) | DMD · COCO last 40 + first 40 | final mAP: 38.89 -> 39.14 (Euclidean 39.58) | No clear effect |
| multiply semantic and feature components instead of adding (ablation) | DMD and IFD · COCO 70+10 | mAP: 37.50 -> 37.60 | Improved |
| add DMD+IFD to LwF label distillation (ablation) | LwF · COCO 70+10 | mAP: 36.40 -> 37.60 | Improved |
| raise alpha and beta beyond 1.5 (ablation) | Total loss weights · COCO 70+10 | mAP: 37.60 -> 37.10 (alpha 2, beta 1) | Hurt |
| increase number of decoder queries from 100 to 300 (ablation) | Deformable DETR · COCO 70+10 | mAP: 36.20 -> 37.60 | Improved |
| use a better AdaMixer teacher and student (ablation) | Transformer detectors · COCO 70+10 | mAP: 37.60 -> 40.20 | Improved |
Closest papers by meaning
- Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector (ICML 2025 · similarity 0.93)
- High-dimension Prototype is a Better Incremental Object Detection Learner (ICLR 2025 · similarity 0.92)
- Overcoming Catastrophic Forgetting in Incremental Object Detection via Elastic Response Distillation (CVPR 2022 · similarity 0.92)
- DCA: Dividing and Conquering Amnesia in Incremental Object Detection (AAAI 2025 · similarity 0.91)
- Learning Task-Aware Language-Image Representation for Class-Incremental Object Detection (AAAI 2024 · similarity 0.89)
- GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation (AAAI 2025 · similarity 0.88)
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.
