Research map / Continual learning / Class-incremental
Revisiting Generative Replay for Class Incremental Object Detection
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.
Revisits generative replay for class-incremental object detection by using an off-the-shelf Stable Diffusion model to generate single-object image-level replay data, pseudo-labeled by detectors, plus a similarity-based cross-sampling step.
Problem
Class-incremental object detection (CIOD) suffers catastrophic forgetting, and generative replay is hard to apply because diffusion models struggle to generate complex multi-object images with precise spatial layouts.
Approach
Observes that forgetting in CIOD mainly affects classification rather than localization, so it uses standard Stable Diffusion to generate image-level (mostly single-object) replay images, assigns boxes via pseudo-labeling from the old and a stage-wise detector, and selects confusing samples with Similarity-based Cross Sampling based on IoU between old and new detectors' predictions.
Key results
- On VOC 10-10 single-step setting, RGR achieves 75.8 mAP, improving over BPF by 2.9%.
- On MS COCO 2017 40+40, RGR reaches 35.6 AP, outperforming BPF (34.4 AP) by 1.2%.
- In the VOC 10-5 (3 tasks) multi-step setting, RGR reaches 71.8 mAP, a 3.1% improvement over the prior best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add IGR image-level generative replay (ablation) | Pseudo-labeling baseline detector · PASCAL VOC 2007 10-10 | mAP: 73.7 -> 75.1 | Helped |
| Add IGR image-level generative replay (ablation) | Pseudo-labeling baseline detector · PASCAL VOC 2007 15-5 | mAP: 67.3 -> 73.8 | Helped |
| Add IGR image-level generative replay (ablation) | Pseudo-labeling baseline detector · PASCAL VOC 2007 10-5 | mAP: 60.4 -> 71.2 | Helped |
| Add SCS to IGR (ablation) | IGR detector · PASCAL VOC 2007 10-10 | mAP: 75.1 -> 75.8 | Helped |
| Add SCS to IGR (ablation) | IGR detector · PASCAL VOC 2007 15-5 | mAP: 73.8 -> 74.0 | Helped |
| Add SCS to IGR (ablation) | IGR detector · PASCAL VOC 2007 10-5 | mAP: 71.2 -> 71.8 | Helped |
| Increase IGR confidence threshold τ from 0.50 to 0.90 (ablation) | IGR detector · PASCAL VOC 2007 10-10 | mAP: 75.1 -> 75.8 | Helped |
| Increase IGR confidence threshold τ from 0.90 to 0.95 (ablation) | IGR detector · PASCAL VOC 2007 10-10 | mAP: 75.8 -> 75.4 | Hurt |
| Use RGR rather than BPF | CIOD detector · MS COCO 2017 40+40 | AP: +1.2% | Helped |
| Use RGR rather than BPF | CIOD detector · MS COCO 2017 40+40 | AP50: +1.7% | Helped |
| Use RGR rather than BPF | CIOD detector · MS COCO 2017 40+40 | AP75: +0.9% | Helped |
| Use RGR rather than BPF | CIOD detector · MS COCO 2017 70+10 | AP50: 56.8 -> 56.6 | Hurt |
Surprising findings
- Old-model pseudo-labels often localize objects correctly while assigning wrong classes, so the authors filter pseudo-labels that conflict with new-task ground truth.
- Forgetting in CIOD is concentrated in classification: class-agnostic box regression accuracy (oAP) stays high even when per-class AP drops to 0.
Limitations
- Generated images are mostly single-object, so complex multi-object scenes are not modeled.
- Synthetic-to-real domain gap remains; loss weighting for synthetic data is a hand-tuned hyperparameter.
- Hyperparameters tau and eta are tuned on the evaluation settings.
Experimental setup
Datasets: PASCAL VOC 2007 MS COCO 2017 synthetic images from Stable Diffusion 1.5
Benchmarks: PASCAL VOC 2007 MS COCO 2017
Metrics: mAP@0.5 AP AP50 AP75
Baselines: Fine-tuning Joint Training ORE ILOD-Meta OW-DETR ABR Faster ILOD PPAS MVC MMA PseudoRM PROB
Builds on: BPF Faster R-CNN SDDGR
Closest papers by meaning
- SDDGR: Stable Diffusion-based Deep Generative Replay for Class Incremental Object Detection (CVPR 2024 · similarity 0.94)
- Demystifying Catastrophic Forgetting in Two-Stage Incremental Object Detector (ICML 2025 · similarity 0.90)
- Learning Task-Aware Language-Image Representation for Class-Incremental Object Detection (AAAI 2024 · similarity 0.89)
- Autoencoder-Based Hybrid Replay for Class-Incremental Learning (ICML 2025 · similarity 0.88)
- Task Confusion and Catastrophic Forgetting in Class-Incremental Learning: A Mathematical Framework for Discriminative and Generative Modelings (NeurIPS 2024 · similarity 0.88)
- Tensor Decomposition Based Memory-Efficient Incremental Learning (ICML 2025 · 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.