Research map / Continual learning / Class-incremental

DCA: Dividing and Conquering Amnesia in Incremental Object Detection

AAAI 2025 · Continual learning › Class-incremental · approach: Incremental object detection with replay · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DCA, which decouples DETR-based incremental object detection into a localization-then-recognition process and uses PLM class-name semantics (duplex classifier fusion, semantic queries) to reduce recognition forgetting, achieving exemplar-free SOTA on VOC and COCO.

Problem

Catastrophic forgetting in transformer-based (DETR) incremental object detection, where recognition forgets severely while localization is comparatively robust; existing methods focus on KD and exemplar replay without addressing the intrinsic forgetting mechanism.

Approach

Identify a forgetting imbalance between localization (less-forgetting, class-agnostic) and recognition (fragile); decouple the decoder into a localization decoder and a semantic-guided recognition decoder, feeding PLM class-name embeddings as queries and adding a semantic-similarity classifier fused with a linear head to anchor class representations across tasks.

Key results

• On VOC 10+10, DCA reaches 73.4 mAP, improving over the best non-exemplar method ACF (68.6) by 4.8 points.
• On COCO four-step (40+10 increments), DCA achieves 37.2 AP-final-step-range with AbsGap 5.4 AP versus ACF-style baselines, and the paper reports a 6.9% final AP improvement under the four-step setting.
• Ablation on VOC 15+5 shows the full DCA framework reaches 71.2 mAP, +8.7 over the pseudo-label fine-tuning baseline (62.5).

Measured effects reported in the paper

Surprising findings

• Using only hybrid knowledge distillation yields 66.0 mAP, and localization remains about 94% recall on unseen classes, indicating localization generalizes to future classes without distillation.

Limitations

• Relies on pseudo-labels of old classes for foreground supervision, with distillation from the old model.
• Fusion weight beta fixed at 0.5 without adaptive selection.
• Gains on some classes (e.g., plant, sheep) remain low in the VOC 15+5 setting.

Experimental setup

Datasets: PASCAL VOC MS COCO ImageNet

Benchmarks: PASCAL VOC MS COCO

Metrics: AP AP50 mAP AbsGap RelGap

Baselines: Faster ILOD MMA ORE-EBUI OW-DETR CL-DETR PROB ACF LwF RILOD SID ERD

Builds on: Deformable DETR CL-DETR ACF

Closest papers by meaning

• Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector (ICML 2025 · similarity 0.91)
• YOLO-IOD: Towards Real Time Incremental Object Detection (AAAI 2026 · similarity 0.90)
• Better Matching, Less Forgetting: A Quality-Guided Matcher for Transformer-based Incremental Object Detection (AAAI 2026 · similarity 0.89)
• Learning Task-Aware Language-Image Representation for Class-Incremental Object Detection (AAAI 2024 · similarity 0.88)
• Revisiting Generative Replay for Class Incremental Object Detection (CVPR 2025 · similarity 0.87)
• GCD: Advancing Vision-Language Models for Incremental Object Detection via Global Alignment and Correspondence Distillation (AAAI 2025 · similarity 0.86)