Research map / Detection and segmentation / Open-set and few-shot detection
Towards Unsupervised Model Selection 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 Detection Adaptation Score (DAS), an unsupervised model selection metric for domain adaptive object detection that combines a flatness index (FIS) and a prototypical distance ratio (PDR) to pick checkpoints without target labels.
Problem
Model selection for domain adaptive object detection (DAOD) usually relies on labeled target data, which is unavailable in real deployments; DAOD training is also unstable and can degrade on the target domain over time.
Approach
Approximate flat-minima generalization without target labels: FIS measures classification and localization fluctuation under random parameter perturbations, and PDR measures source-target class-prototype distance ratios as a proxy for transferability and discriminability; the two are min-max normalized and summed into DAS.
Key results
- DAS-selected checkpoints average 42.52% mAP versus 39.27% for last checkpoints and 43.39% oracle across four DAOD methods on three benchmarks.
- DAS achieves a PCC of 0.86 between its score and ground-truth mAP across checkpoints, versus poor correlation for prior unsupervised metrics.
- On the hyperparameter-tuning task, DAS reaches PCC 0.865 (lambda_dis) and 0.938 (lambda_unsup) on Weather adaptation, outperforming baselines such as FD (0.631 and 0.751).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select a checkpoint using DAS instead of taking the last checkpoint | DAOD models · Real-to-Art (P2C), Weather Adaptation (C2F), and Synthetic-to-Real (S2C) | mAP: +5.85%, +1.1%, and +19.09% | Helped |
| select checkpoints using DAS rather than the last checkpoint | DAOD models, averaged across benchmarks and frameworks · Three DAOD benchmarks and four frameworks | mAP: 39.27 -> 42.52 | Helped |
| select checkpoints using DAS instead of TS | DAOD models, averaged across benchmarks and frameworks · Three DAOD benchmarks and four frameworks | mAP: $2.42\%$ | Helped |
| select checkpoints using DAS instead of BoS | DAOD models, averaged across benchmarks and frameworks · Three DAOD benchmarks and four frameworks | mAP: $3.60\%$ | Helped |
| add FIS to a PDR-only score (ablation) | DAS · Three DAOD benchmarks and four frameworks | mAP: 42.14 -> 42.52 | Helped |
| add PDR to an FIS-only score (ablation) | DAS · Three DAOD benchmarks and four frameworks | mAP: 41.74 -> 42.52 | Helped |
| add FIS to a PDR-only score (ablation) | DAS · Three DAOD benchmarks and four frameworks | PCC: 0.64 -> 0.67 | Helped |
| add PDR to an FIS-only score (ablation) | DAS · Three DAOD benchmarks and four frameworks | PCC: 0.48 -> 0.67 | Helped |
| set lambda to 0.1 rather than the last checkpoint (ablation) | DAS · Real-to-Art (P2C) | mAP: 33.02 -> 34.76 | Helped |
| set lambda to 0.5 rather than the last checkpoint (ablation) | DAS · Real-to-Art (P2C) | mAP: 33.02 -> 35.08 | Helped |
| set lambda to 1.0 rather than the last checkpoint (ablation) | DAS · Real-to-Art (P2C) | mAP: 33.02 -> 35.09 | Helped |
| set lambda to 2.0 rather than the last checkpoint (ablation) | DAS · Real-to-Art (P2C) | mAP: 33.02 -> 35.05 | Helped |
Surprising findings
- Adaptive Teacher on Sim10k-to-Real drops to 28.35% at the last checkpoint versus 47.44% for DAS-selected checkpoint, showing severe late-training degradation.
- The BoS baseline underperforms the last-checkpoint baseline on average (38.92 vs 39.27), indicating existing detection-specific unsupervised metrics can fail under domain shift.
Limitations
- Requires the entire source and target datasets to compute the score, which is data-inefficient.
- Does not consistently select the best checkpoint; a notable gap to oracle remains on some settings (e.g., CMT on S2C: 48.81 vs 55.53).
- Hyperparameters lambda and gamma were fixed to 1 across experiments without per-benchmark tuning; correlation is weaker in some settings.
Experimental setup
Datasets: PASCAL VOC 2007/2012 Clipart1k Cityscapes Foggy Cityscapes Sim10k
Benchmarks: Clipart1k (Real-to-Art P2C) Foggy Cityscapes (Weather C2F) Cityscapes (Sim10k-to-Real S2C)
Metrics: mAP@0.5 Pearson correlation coefficient (PCC)
Baselines: Prediction Score (PS) Entropy Score (ES) Average Threshold Confidence (ATC) Fréchet Distance (FD) Transfer Score (TS) Bounding Box Stability (BoS)
Builds on: Adaptive Teacher Mean Teacher DA Faster RCNN Transfer Score (TS) Bounding Box Stability (BoS)
Closest papers by meaning
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.89)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.88)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.88)
- Triple Feature Disentanglement for One-Stage Adaptive Object Detection (AAAI 2024 · similarity 0.87)
- Non-exemplar Domain Incremental Object Detection via Learning Domain Bias (AAAI 2024 · similarity 0.87)
- Differential Alignment for Domain Adaptive Object Detection (AAAI 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.