Research map / Detection and segmentation / Open-set and few-shot detection
DSD-DA: Distillation-based Source Debiasing 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 DSD-DA, which distills domain-agnostic knowledge from a mix-style classification teacher into DA-Faster and refines classification scores at test time with a target-relevant localization score from TROLN, improving cross-domain detection mAP.
Problem
Feature-alignment domain adaptive object detectors inherit source bias (more source-specific knowledge) and suffer exacerbated inconsistency between classification and localization on the target domain.
Approach
Distill domain-agnostic knowledge into the detector from a classification teacher trained on source and CycleGAN source-to-target-style images, and build a target-affinity-weighted localization score (from a TROLN localization teacher using centerness and IoU) that refines classification scores in testing and NMS.
Key results
- On Cityscapes to FoggyCityscapes with DA-Faster (ResNet50), DSD-DA reaches 45.2 mAP vs 40.2 for DA-Faster, and with VGG16 reaches 44.2 mAP vs 39.6.
- On KITTI to Cityscapes with VGG-16, DA-Faster + DSD-DA reaches 46.9 car AP50 vs 43.4 for DA-Faster; AT + DSD-DA reaches 49.3 vs 47.7 for AT.
- Ablation on FoggyCityscapes with VGG16 shows DSD alone raises mAP from 39.56 to 42.00 and DCE alone to 42.14, and combined reach 44.21.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DSD-DA to DA-Faster | DA-Faster with VGG16 · Cityscapes→FoggyCityscapes | mAP: 39.6 -> 44.2 | Helped |
| Add DSD-DA to AT | AT with VGG16 · Cityscapes→FoggyCityscapes | mAP: 49.3 -> 52.2 | Helped |
| Add DSD-DA to DA-Faster | DA-Faster with VGG16 · Kitti→Cityscapes | AP50 (Car): 43.4 -> 46.9 | Helped |
| Add DSD-DA to DA-Faster | DA-Faster · SIM10k→Cityscapes | AP50 (Car): 43.6 -> 47.8 | Helped |
| Add DSD without DCE (ablation) | DA-Faster with VGG16 · FoggyCityscapes validation | mAP: 39.56 -> 42.00 | Helped |
| Add DCE without DSD (ablation) | DA-Faster with VGG16 · FoggyCityscapes validation | mAP: 39.56 -> 42.14 | Helped |
| Add both DSD and DCE (ablation) | DA-Faster with ResNet50 · FoggyCityscapes validation | mAP: 40.15 -> 45.22 | Helped |
| Add target samples filtered by objectness to source distillation (ablation) | DSD, DA-Faster ResNet50 · FoggyCityscapes test | AP50: 43.05 -> 38.87 | Hurt |
| Add target samples filtered by classification to source distillation (ablation) | DSD, DA-Faster ResNet50 · FoggyCityscapes test | AP50: 43.05 -> 37.64 | Hurt |
| Add target samples filtered by objectness and classification to source distillation (ablation) | DSD, DA-Faster ResNet50 · FoggyCityscapes test | AP50: 43.05 -> 38.43 | Hurt |
| Use centerness as the localization representation (ablation) | OLN with DA-Faster ResNet50 · FoggyCityscapes test | AP50: 43.05 -> 43.95 | Helped |
| Use IoU as the localization representation (ablation) | OLN with DA-Faster ResNet50 · FoggyCityscapes test | AP50: 43.05 -> 43.98 | Helped |
Surprising findings
- Distilling target-style or pseudo-labeled target samples into the detector decreased performance, so only source samples were used for distillation.
- Training the localization network on only CycleGAN-generated target-style data did not effectively capture target-relevant information (AP50 35.76 vs 43.05 baseline with source and T' combined).
Limitations
- Relies on CycleGAN-generated source-to-target style images, whose domain shift from real target data limits localization learning.
- Distillation with target data was harmful due to noisy pseudo-labels, so the method is restricted to source-only distillation.
- Reported gains are measured on a small set of benchmarks with AP50 on car only for two of three settings.
Experimental setup
Datasets: Cityscapes FoggyCityscapes KITTI SIM10k ImageNet
Benchmarks: Cityscapes to FoggyCityscapes KITTI to Cityscapes SIM10k to Cityscapes
Metrics: mAP AP50 AP_s AP_t Spearman rank correlation Kendall rank correlation TIDE error analysis proxy A-distance
Baselines: DA-Faster AT CMT SWDA TIA TDD MGA PT SCAN SIGMA SIGMA++ HT
Builds on: DA-Faster OLN CycleGAN Faster R-CNN GRL
Closest papers by meaning
- Differential Alignment for Domain Adaptive Object Detection (AAAI 2025 · similarity 0.90)
- DA-Ada: Learning Domain-Aware Adapter for Domain Adaptive Object Detection (NeurIPS 2024 · similarity 0.89)
- Universal Domain Adaptive Object Detection via Dual Probabilistic Alignment (AAAI 2025 · similarity 0.89)
- Improving Single Domain-Generalized Object Detection: A Focus on Diversification and Alignment (CVPR 2024 · similarity 0.89)
- Triple Feature Disentanglement for One-Stage Adaptive Object Detection (AAAI 2024 · similarity 0.88)
- Towards Unsupervised Model Selection for Domain Adaptive Object Detection (NeurIPS 2024 · 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.