Research map / Label-efficient and robust learning / Domain adaptation and generalization
Train Till You Drop: Towards Stable and Robust Source-free Unsupervised 3D Domain Adaptation
Björn Michele, Alexandre Boulch, Tuan-Hung VU, Gilles Puy, Renaud Marlet, Nicolas Courty
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add TTYD_core (entropy + source-class-distribution loss, BN-to-linear adaptation, agreement-based stopping) vs best hyperparameter-free baseline | Minkowski U-Net SFUDA · nuScenes->SemanticKITTI | mIoU: 41.7 -> 44.5 | Improved |
| add self-training module on top of TTYD_core | TTYD_core · nuScenes->SemanticKITTI | mIoU: 44.5 -> 45.4 | Improved |
| use agreement-with-reference stopping criterion instead of entropy validator (ablation) | TTYD training scheme · nuScenes->SemanticPOSS | mIoU: 23.7 -> 61.1 | Improved |
| add stopping criterion to training scheme that otherwise degrades (ablation) | TTYD training scheme · nuScenes->SemanticPOSS | mIoU: -38.0 pp drop without stopping | Improved |
| add source-class-distribution loss (ablation) | TTYD training scheme · nuScenes->SemanticKITTI | max mIoU: 34.4 -> 40.9 | Improved |
| add entropy discriminability loss alongside distribution loss (ablation) | TTYD training scheme · nuScenes->SemanticKITTI | max mIoU: 40.9 -> 44.7 | Improved |
| add entropy loss alone, or uniform-distribution loss alone (ablation) | TTYD training scheme · nuScenes->SemanticKITTI | max mIoU: 34.4 -> 34.4 | No clear effect |
| use source class distribution instead of uniform as prior (ablation) | TTYD training scheme · nuScenes->SemanticKITTI | max mIoU: 35.0 -> 44.7 | Improved |
| adapt only the classification layer (ablation) | TTYD training scheme · nuScenes->SemanticKITTI | max mIoU: 44.7 -> 34.4 | Hurt |
| adapt BN running statistics online at train and eval time (ablation) | TTYD training scheme · nuScenes->SemanticKITTI | max mIoU: 44.7 -> 39.1 | Hurt |
| use MeanBN rather than PTBN as the reference model for stopping (ablation) | TTYD_stop · nuScenes->SemanticKITTI | mIoU: 44.5 -> 39.0 | Hurt |
| use L1, L2 or symmetric KL instead of hard agreement counting (ablation) | TTYD_stop · nuScenes->SemanticKITTI | mIoU: 44.5 -> 44.5 | No clear effect |
Closest papers by meaning
- Source-Free Domain Adaptation for Semantic Segmentation (CVPR 2021 · similarity 0.90)
- UniDSeg: Unified Cross-Domain 3D Semantic Segmentation via Visual Foundation Models Prior (NeurIPS 2024 · similarity 0.89)
- Universal Domain Adaptation for Semantic Segmentation (CVPR 2025 · similarity 0.88)
- Open-Set Domain Adaptation for Semantic Segmentation (CVPR 2024 · similarity 0.88)
- Stable Neighbor Denoising for Source-free Domain Adaptive Segmentation (CVPR 2024 · similarity 0.88)
- Omni-Query Active Learning for Source-Free Domain Adaptive Cross-Modality 3D Semantic Segmentation (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.
