Research map / Continual learning / Class-incremental
Incremental Learning of Object Detectors Without Catastrophic Forgetting
Konstantin Shmelkov, Cordelia Schmid, Karteek Alahari
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 L2 distillation loss on old-class logits (ablation) | B(20) incremental detector · VOC 2007 test (add 1 class tvmonitor) | mAP old classes: 25.0 -> 68.3 | Improved |
| add L2 distillation loss on old-class logits (ablation) | B(11-20) incremental detector · VOC 2007 test (add 10 classes) | mAP old classes: 12.8 -> 63.2 | Improved |
| finetune the conv trunk instead of freezing it (ablation) | B(20) with distillation · VOC 2007 test (add 1 class tvmonitor) | mAP new class: 43.2 -> 58.3 | Improved |
| add bounding box regression distillation (ablation) | B(11-20) · VOC 2007 test (add 10 classes) | mAP all classes: 60.9 -> 63.1 | Improved |
| add bounding box regression distillation (ablation) | B(20) · VOC 2007 test (add 1 class tvmonitor) | mAP new class: 62.7 -> 58.3 | Hurt |
| use L2 distillation instead of cross-entropy distillation (ablation) | B(20) · VOC 2007 test (add 1 class tvmonitor) | mAP new class: 52.0 -> 58.3 | Improved |
| use L2 distillation instead of cross-entropy distillation (ablation) | B(20) · VOC 2007 test (add 1 class tvmonitor) | mAP old classes: 67.3 -> 67.8 | No clear effect |
| sample non-background (lowest background score) proposals instead of random ones for distillation (ablation) | B(16)...(20) sequential · VOC 2007 test (add 5 classes) | mAP all classes: 46.0 -> 62.4 | Improved |
| add all new classes at once instead of sequentially (ablation) | B(16-20) · VOC 2007 test (add 5 classes) | mAP all classes: 62.4 -> 65.9 | Improved |
| use distillation instead of EWC (ablation) | B(11-20) · VOC 2007 test (add 10 classes) | mAP all classes: 46.3 -> 63.1 | Improved |
| train second stage on COCO images instead of VOC images (ablation) | B(11-20) · VOC 2007 test (add 10 classes, second stage data) | mAP all classes: 63.2 -> 54.8 | Hurt |
| train one separate network per new class instead of one incremental network (ablation) | A(1-15)+5 classes · VOC 2007 test (add 5 classes) | mAP all classes: 65.9 -> 62.4 | Hurt |
Closest papers by meaning
- Bridging Non Co-occurrence with Unlabeled In-the-wild Data for Incremental Object Detection (NeurIPS 2021 · similarity 0.92)
- Towards Non Co-occurrence Incremental Object Detection with Unlabeled In-the-Wild Data (IJCV 2024 · similarity 0.91)
- Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector (ICML 2025 · similarity 0.90)
- Revisiting Generative Replay for Class Incremental Object Detection (CVPR 2025 · similarity 0.89)
- DCA: Dividing and Conquering Amnesia in Incremental Object Detection (AAAI 2025 · similarity 0.88)
- Learning Task-Aware Language-Image Representation for Class-Incremental Object Detection (AAAI 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.
