Research map / Detection and segmentation / Open-set and few-shot detection
Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment
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 Dual Probabilistic Alignment (DPA), a Faster-RCNN-based universal domain adaptive object detection method that models domain probabilities as Gaussians to align domain-private categories globally and domain-shared categories at the instance level, plus a private class constraint.
Problem
Universal domain adaptive object detection (UniDAOD) must handle open-set, partial-set, and closed-set shifts; existing methods align only shared categories at global and instance levels, ignoring domain-private alignment at the global level and suffering from heterogeneous domain probabilities acro
Approach
Model domain discriminator probabilities as Gaussian distributions; use CDF-based weights for global-level domain-private alignment (GDPA), Gaussian-binned gradient-norm sampling and weights for instance-level domain-shared alignment (IDSA), and a cross-space centroid consistency constraint (PCC) to reduce negative transfer.
Key results
- DPA reaches 41.4% mAP on Pascal VOC to Clipart1k open-set beta=75%, above CODE's 39.4%.
- DPA reaches 42.7% mAP on open-set beta=50% and 42.5% at beta=25%, improving over the best reported UniDAOD baselines.
- On Cityscapes to Foggy Cityscapes closed-set, DPA reports 46.3% mAP versus 42.1% for CODE and 44.5% for SIGMA++.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DPA instead of CODE | UniDAOD on Pascal VOC to Clipart1k · Open-set β=75% | mAP: 39.4 -> 41.4 | Helped |
| Use DPA instead of CODE | UniDAOD on Pascal VOC to Clipart1k · Open-set β=50% | mAP: 40.6 -> 42.7 | Helped |
| Use DPA instead of CODE | UniDAOD on Pascal VOC to Clipart1k · Open-set β=25% | mAP: 41.2 -> 42.5 | Helped |
| Use DPA instead of US-DAF | UniDAOD on WaterColor to Pascal VOC · Partial-set Cs⊂Ct | mAP: 52.9 -> 54.1 | Helped |
| Use DPA instead of CODE | UniDAOD on Pascal VOC to WaterColor · Partial-set Cs⊃Ct | mAP: 55.8 -> 56.3 | Helped |
| Use DPA instead of SIGMA++ | Object detection on Cityscapes to Foggy Cityscapes · Closed-set | mAP: 44.5 -> 46.3 | Helped |
| Add GDPA (ablation) | DPA on Pascal VOC to Clipart1k · Open-set β=75% | mAP: 38.3 -> 41.4 | Helped |
| Add GDPA (ablation) | DPA on Pascal VOC to Clipart1k · Open-set β=50% | mAP: 39.7 -> 42.7 | Helped |
| Add GDPA (ablation) | DPA on Pascal VOC to Clipart1k · Open-set β=25% | mAP: 40.1 -> 42.5 | Helped |
| Add IDSA (ablation) | DPA on Pascal VOC to Clipart1k · Open-set β=75% | mAP: 36.3 -> 41.4 | Helped |
| Add IDSA (ablation) | DPA on Pascal VOC to Clipart1k · Open-set β=50% | mAP: 40.1 -> 42.7 | Helped |
| Add IDSA (ablation) | DPA on Pascal VOC to Clipart1k · Open-set β=25% | mAP: 39.9 -> 42.5 | Helped |
Surprising findings
- Removing the PCC module lowers performance at beta=25% (39.4 vs 42.5) while removing GDPA/IDSA hurts more at high shared ratios, reversing the expected pattern per module.
Limitations
- Gradient-norm bin size and several hyperparameters (delta, alpha, gamma) are set manually or by the authors without detailed sensitivity analysis.
- Theoretical analysis is loosely derived from UDA bounds and relies on assumed symmetric hypothesis space.
- Closed-set gains are reported only on one benchmark pair.
Experimental setup
Datasets: Pascal VOC Clipart1k Watercolor Cityscapes Foggy Cityscapes
Benchmarks: Pascal VOC to Clipart1k Pascal VOC to Watercolor Cityscapes to Foggy Cityscapes
Metrics: mAP@0.5
Baselines: Source Only DAF MAF HTCN UAN CMU SFA US-DAF UCF CODE SCAN++ SIGMA++
Builds on: Faster R-CNN DAF CODE US-DAF focal loss
Closest papers by meaning
- Differential Alignment for Domain Adaptive Object Detection (AAAI 2025 · similarity 0.93)
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.92)
- Large Self-Supervised Models Bridge the Gap in Domain Adaptive Object Detection (CVPR 2025 · similarity 0.91)
- SEEN-DA: SEmantic ENtropy guided Domain-aware Attention for Domain Adaptive Object Detection (CVPR 2025 · similarity 0.90)
- Mean Teacher DETR with Masked Feature Alignment: A Robust Domain Adaptive Detection Transformer Framework (AAAI 2024 · similarity 0.89)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.89)
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.