Research map / Detection and segmentation / Open-set and few-shot detection
Active Domain Adaptation with False Negative Prediction for 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 the first active domain adaptation method for object detection, using a False Negative Prediction Module to score undetectability alongside localization uncertainty, entropy, and diversity for sample selection, plus uncertainty-guided pseudo-labeling in a mean-teacher framework.
Problem
Under domain shift, object detectors suffer from false negative errors on undetected objects, but existing active sampling methods estimate uncertainty only from predicted boxes and so cannot account for undetected objects.
Approach
A FNPM branch on the domain-adapted backbone regresses the number of false negatives per image; this undetectability score is multiplied with localization uncertainty (MC dropout), classification entropy, and a discriminator-based diversity term to select target images, and low-variance pseudo-labels are used for semi-supervised training.
Key results
- With 1% labeling budget, K→C reaches 64.2 AP versus 60.2 for PT (UDA) and 68.9 for Oracle; with 5% budget, C→B reaches 50.9 vs Oracle 52.3.
- Ours (5%) reaches 98% accuracy difference ratio relative to Oracle, nearly matching fully supervised training.
- On C→B at 0.5% budget, the method gains 2.5 pt over AADA; the undetectability metric sfn improves AADA and MeanEntropy at low budgets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FNPM-based undetectability to active sampling (ablation) | MeanEntropy sampling · C → B | mAP: 40.5 -> 41.7; 44.7 -> 43.9; 49.2 -> 48.4 | Mixed |
| add FNPM-based undetectability to active sampling (ablation) | AADA sampling · C → B | mAP: 35.8 -> 37.4; 39.5 -> 38.3; 40.1 -> 40.7; 43.5 -> 44.0; 48.0 -> 48.4; 49.4 -> 50.4 | Mixed |
| add strong augmentation (ablation) | Full method · C → F, 0.5% budget | AP: 38.2 -> 51.6 | Helped |
| add strong augmentation (ablation) | Full method · K → C, 0.5% budget | AP: 63.2 -> 63.1 | No clear effect |
| add adversarial learning, mean teacher, and uncertainty-aware pseudo-labeling (ablation) | PL + SA · C → F, 0.5% budget | AP: 40.1 -> 51.6 | Helped |
| add adversarial learning, mean teacher, and uncertainty-aware pseudo-labeling (ablation) | PL + SA · K → C, 0.5% budget | AP: 53.9 -> 63.1 | Helped |
| add uncertainty-aware pseudo-labeling (ablation) | AD + PL + MT + SA · C → F, 0.5% budget | AP: 51.1 -> 51.6 | No clear effect |
| add uncertainty-aware pseudo-labeling (ablation) | AD + PL + MT + SA · K → C, 0.5% budget | AP: 54.2 -> 63.1 | Helped |
| add mean teacher (ablation) | AD + PL + UP + SA · C → F, 0.5% budget | AP: 50.5 -> 51.6 | Helped |
| add mean teacher (ablation) | AD + PL + UP + SA · K → C, 0.5% budget | AP: 60.9 -> 63.1 | Helped |
| add adversarial learning (ablation) | PL + MT + UP + SA · C → F, 0.5% budget | AP: 30.2 -> 51.6 | Helped |
| add adversarial learning (ablation) | PL + MT + UP + SA · K → C, 0.5% budget | AP: 49.9 -> 63.1 | Helped |
Surprising findings
- Analysis shows FN errors, not FP errors, dominate the performance gap under domain shift even after UDA.
- Mean-teacher and uncertainty-aware pseudo-labeling contributed little on C→F, while adversarial alignment contributed little on K→C.
Limitations
- Evaluation limited to in-vehicle driving scenarios and a single detector (Faster R-CNN with VGG16).
- Reproduced AADA themselves (AADA†) rather than using official results.
- Sampling relies on several sequential stages (UDA init, FNPM training, alternating updates), adding complexity.
Experimental setup
Datasets: Cityscapes Foggy Cityscapes BDD100k SIM10k KITTI
Benchmarks: Cityscapes to Foggy Cityscapes Cityscapes to BDD100k SIM10k to Cityscapes KITTI to Cityscapes
Metrics: AP@0.5 mAP@0.5
Baselines: Source-only DA-Faster SWDA PT AT MGADA CMT CSDA NSA-UDA AADA Random MeanEntropy
Builds on: Faster R-CNN Probabilistic Teacher AADA MC Dropout Mean Teacher
Closest papers by meaning
- CAT: Exploiting Inter-Class Dynamics for Domain Adaptive Object Detection (CVPR 2024 · similarity 0.87)
- Plug and Play Active Learning for Object Detection (CVPR 2024 · similarity 0.87)
- Category-Aware Active Domain Adaptation (ICML 2024 · similarity 0.87)
- Composite Active Learning: Towards Multi-Domain Active Learning with Theoretical Guarantees (AAAI 2024 · similarity 0.86)
- Revisiting the Domain Shift and Sample Uncertainty in Multi-source Active Domain Transfer (CVPR 2024 · similarity 0.86)
- Towards Unsupervised Model Selection for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.86)
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.