Research map / Label-efficient and robust learning / Semi-supervised learning
Enhancing Generalizability via Utilization of Unlabeled Data for Occupancy Perception
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 UGOCC, a domain-adaptation and semi-supervised framework that uses unlabeled target-domain driving data to improve camera-only 3D occupancy perception generalization between nuScenes and Waymo.
Problem
Vision-based 3D occupancy perception models degrade sharply when transferred across driving datasets with different sensor suites and scenes, and dense occupancy labels are expensive.
Approach
Combines semantic-query adversarial fusion (semantic attention as a learnable query with domain adversarial alignment), self-supervised multi-view depth augmentation on unlabeled temporal data, and an occupancy-specific pseudo-label selection using confidence, spatial smoothness, and 2D-to-3D rendering consistency.
Key results
- On the nuScenes-to-Waymo transition, UGOCC reaches 16.7 target mIoU versus 1.9-4.3 for baselines (BEVDet*, BEVF*, FBOCC*, Flash*).
- On the Waymo-to-nuScenes transition, UGOCC reaches 16.8 target mIoU versus 0.8-4.2 for baselines.
- UGOCC keeps source-domain mIoU comparable to baselines (32.0 vs 30.2-32.5 on nuScenes to Waymo; 23.3 on Waymo to nuScenes).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply UGOCC using unlabeled target-domain data | occupancy perception · nuScenes to Waymo (target) | mIoU: 3.9 -> 16.7 | Helped |
| Apply UGOCC using unlabeled target-domain data | occupancy perception · Waymo to nuScenes (target) | mIoU: 4.2 -> 16.8 | Helped |
| Add Semantic Query Adversarial Fusion (ablation) | UGOCC · nuScenes to Waymo (target) | mIoU: 15.2 -> 16.7 | Helped |
| Add Semantic Query Adversarial Fusion (ablation) | UGOCC · Waymo to nuScenes (target) | mIoU: 14.9 -> 16.8 | Helped |
| Add self-supervised depth augmentation (ablation) | UGOCC · nuScenes to Waymo (target) | mIoU: 13.1 -> 16.7 | Helped |
| Add self-supervised depth augmentation (ablation) | UGOCC · Waymo to nuScenes (target) | mIoU: 15.3 -> 16.8 | Helped |
| Add self-supervised depth augmentation and Semantic Query Adversarial Fusion (ablation) | UGOCC · nuScenes to Waymo (target) | mIoU: 11.8 -> 16.7 | Helped |
| Add self-supervised depth augmentation and Semantic Query Adversarial Fusion (ablation) | UGOCC · Waymo to nuScenes (target) | mIoU: 13.1 -> 16.8 | Helped |
| Add Semantic Query Adversarial Fusion (ablation) | UGOCC · nuScenes to Waymo (source) | mIoU: 32.3 -> 32.0 | Hurt |
| Add self-supervised depth augmentation (ablation) | UGOCC · Waymo to nuScenes (source) | mIoU: 19.1 -> 23.3 | Helped |
Surprising findings
- Baseline occupancy models collapse to target mIoU below 5 across both transfer directions, with Waymo-source transfers showing especially low scores for FlashOcc (0.8).
Limitations
- Waymo target mIoU remains low in absolute terms (about 16-17 mIoU), and the text attributes lower Waymo scores partly to camera FOV and evaluation mask.
- Semantic labels for key frames are generated automatically with Grounded-SAM and SAM rather than human annotation, which may introduce label noise.
- Evaluated only on two dataset pairs with ResNet-50 backbone and a fixed set of four baselines.
Experimental setup
Datasets: nuScenes Waymo OCC-nuScenes OCC-Waymo
Benchmarks: OCC-nuScenes OCC-Waymo
Metrics: mIoU
Baselines: BEVDetOCC BEVFormer FB-OCC FlashOcc
Builds on: BEVDet BEVFormer FB-OCC FlashOcc Lift-Splat-Shoot
Closest papers by meaning
- Achieving Speed-Accuracy Balance in Vision-based 3D Occupancy Prediction via Geometric-Semantic Disentanglement (AAAI 2025 · similarity 0.89)
- Semi-Supervised Vision-Centric 3D Occupancy World Model for Autonomous Driving (ICLR 2025 · similarity 0.89)
- UnO: Unsupervised Occupancy Fields for Perception and Forecasting (CVPR 2024 · similarity 0.88)
- COTR: Compact Occupancy TRansformer for Vision-based 3D Occupancy Prediction (CVPR 2024 · similarity 0.87)
- Collaborative Semantic Occupancy Prediction with Hybrid Feature Fusion in Connected Automated Vehicles (CVPR 2024 · similarity 0.87)
- SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction (CVPR 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.