Research map / Label-efficient and robust learning / Domain adaptation and generalization
CAT: Exploiting Inter-Class Dynamics for Domain Adaptive 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.
Proposes Class-Aware Teacher (CAT), a mean-teacher domain adaptive object detector that uses an inter-class relation confusion matrix to drive class-relation MixUp augmentation and a class-relation weighted classification loss.
Problem
Domain adaptive object detection with student-teacher frameworks suffers from class imbalance among foreground classes, leading to inaccurate pseudo-labels and poor performance on minority classes.
Approach
An Inter-Class Relation module (ICRm) builds an EMA-updated, ground-truth-normalised confusion matrix to estimate which minority classes are confused with which majority classes; this guides instance-level MixUp between related classes (within and across domains) and a weighted classification loss emphasising minority classes misclassified as majority ones.
Key results
- On Cityscapes to Foggy Cityscapes (0.02), CAT reaches 52.5 mAP versus 51.2 for the prior state of the art MRT and 49.3 for AT.
- On PASCAL VOC to Clipart1K, CAT achieves 49.1 mAP, improving on CMT (47.0) by 2.1 mAP.
- Ablation on Foggy Cityscapes: Inter-Class Loss alone gives 51.6 mAP and ICL plus CRA gives 52.5 mAP, with class-mAP std reduced from 10.8 to 8.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CAT rather than MRT | Faster R-CNN domain-adaptive object detection · Cityscapes → Foggy Cityscapes (0.02) test | mAP: +1.3 mAP | Helped |
| Apply CAT rather than CMT | Domain-adaptive object detection · PASCAL VOC → Clipart1K test | mAP: 2.1 mAP | Helped |
| Add CRA to the base AT framework (ablation) | AT on Cityscapes → Foggy Cityscapes · Cityscapes → Foggy Cityscapes (0.02) | mAP: 49.3 -> 51.0 | Helped |
| Add ICL to the base AT framework (ablation) | AT on Cityscapes → Foggy Cityscapes · Cityscapes → Foggy Cityscapes (0.02) | mAP: 49.3 -> 51.6 | Helped |
| Use both CRA and ICL rather than the base AT framework (ablation) | AT on Cityscapes → Foggy Cityscapes · Cityscapes → Foggy Cityscapes (0.02) | mAP: 49.3 -> 52.5 | Helped |
| Select instance pairs randomly for MixUp (ablation) | CAT augmentation on Cityscapes → Foggy Cityscapes · Cityscapes → Foggy Cityscapes (0.02) | mAP: +0.2 mAP | Helped |
| Use CRA to pair highly related classes rather than selecting pairs randomly (ablation) | MixUp augmentation on Cityscapes → Foggy Cityscapes · Cityscapes → Foggy Cityscapes (0.02) | mAP: +0.9 mAP | Helped |
| Increase CRA augmentation likelihood from 0.5 to 1.0 (ablation) | CAT on Cityscapes → Foggy Cityscapes · Cityscapes → Foggy Cityscapes (0.02) | mAP: 52.5 -> 50.2 | Hurt |
| Use inter-class relationships in ICL rather than class-level diagonal weights (ablation) | Classification loss on Cityscapes → Foggy Cityscapes · Cityscapes → Foggy Cityscapes (0.02) | mAP: +0.5 mAP | Helped |
| Remove the regularisation term from ICL (ablation) | Inter-Class Loss on Cityscapes → Foggy Cityscapes · Cityscapes → Foggy Cityscapes (0.02) | mAP: 52.5 -> 51.3 | Hurt |
Surprising findings
- Class-level diagonal-only weighting performed nearly as well as the full Inter-Class Loss (52.0 vs 52.5 mAP), and removing the regularisation term lowered ICL to 51.3 mAP.
- Applying MixUp augmentation with a high augmentation rate (1.0) hurt performance (50.2 mAP), below random MixUp at 0.5 (51.8 mAP).
Limitations
- Gains over prior methods are modest (about 1.3 mAP on Foggy Cityscapes), and results are from single runs with no variance reported beyond class-mAP std.
- Sim10K and KITTI benchmarks were excluded because they focus only on the car class, limiting evaluation of the class-imbalance setting.
- Augmentation ratio is sensitive; excessive augmentation degrades performance.
Experimental setup
Datasets: Cityscapes Foggy Cityscapes PASCAL VOC 2012 Clipart1K
Benchmarks: Cityscapes to Foggy Cityscapes PASCAL VOC to Clipart1K
Metrics: mAP@0.5 standard deviation of class mAP (σ)
Baselines: AT CMT MRT MILA MeGA TIA UMT TDD PT I3Net ICR-CCR HTCN
Builds on: Adaptive Teacher (AT) mean teacher Faster R-CNN MixUp Cropbank oversampling
Closest papers by meaning
- Cross-Domain Adaptive Teacher for Object Detection (CVPR 2022 · similarity 0.93)
- Exploring Categorical Regularization for Domain Adaptive Object Detection (CVPR 2020 · similarity 0.90)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.89)
- D3T: Distinctive Dual-Domain Teacher Zigzagging Across RGB-Thermal Gap for Domain-Adaptive Object Detection (CVPR 2024 · similarity 0.88)
- Non-exemplar Domain Incremental Object Detection via Learning Domain Bias (AAAI 2024 · similarity 0.88)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (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.