Research map / 3D vision / Depth and stereo
STViT: Improving Self-Supervised Multi-Camera Depth Estimation with Spatial-Temporal Context and Adversarial Geometry Regularization (Student Abstract)
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 STViT, a self-supervised multi-camera depth estimation framework with a Spatial-Temporal Transformer encoder and a GAN-based adversarial geometry regularization module against adverse weather conditions.
Problem
Self-supervised multi-camera depth estimation for autonomous driving struggles with small overlaps between adjacent cameras and adverse conditions such as rain and night, which hurt photometric-consistency training.
Approach
Combines a hybrid local-global Transformer encoder with spatial-temporal cross-attention over overlapping camera/frame views, and regularizes depth distributions with a GAN discriminator that compares predictions against unpaired normal-condition depth maps.
Key results
- On nuScenes at 352x640, STViT reports Abs Rel 0.233 and RMSE 6.681 versus SurroundDepth's 0.245 and 6.835.
- On DDAD at 384x640, STViT reports Abs Rel 0.192 and Sq Rel 2.965, the best values in the table, versus EGA-Depth's 0.195 and 3.211.
- The authors claim state-of-the-art performance on nuScenes and DDAD; the abstract-length student paper gives no ablation results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STViT instead of EGA-Depth | Self-supervised multi-camera depth estimation · nuScenes | Abs Rel: 0.239 -> 0.233 | Helped |
| Use STViT instead of EGA-Depth | Self-supervised multi-camera depth estimation · nuScenes | Sq Rel: 2.357 -> 2.815 | Hurt |
| Use STViT instead of EGA-Depth | Self-supervised multi-camera depth estimation · nuScenes | RMSE: 6.801 -> 6.681 | Helped |
| Use STViT instead of SurroundDepth | Self-supervised multi-camera depth estimation · nuScenes | RMSE log: 0.321 -> 0.312 | Helped |
| Use STViT instead of EGA-Depth | Self-supervised multi-camera depth estimation · DDAD | Abs Rel: 0.195 -> 0.192 | Helped |
| Use STViT instead of EGA-Depth | Self-supervised multi-camera depth estimation · DDAD | Sq Rel: 3.211 -> 2.965 | Helped |
| Use STViT instead of FSM* | Self-supervised multi-camera depth estimation · DDAD | RMSE: 12.270 -> 12.156 | Helped |
| Use STViT instead of EGA-Depth | Self-supervised multi-camera depth estimation · DDAD | RMSE log: 0.297 -> 0.293 | Helped |
Limitations
- Student abstract with limited experimental detail; no ablation study is reported in the text.
- On nuScenes at 352x640, EGA-Depth has the best Sq Rel (2.357) and STViT is not best across all metrics, so the SOTA claim is partial.
- Adverse-condition benefits of the AGR module are asserted but not separately quantified in the provided text.
Experimental setup
Datasets: nuScenes DDAD
Benchmarks: nuScenes DDAD
Metrics: Abs Rel Sq Rel RMSE RMSE log accuracy thresholds
Baselines: Monodepth2 FSM SurroundDepth EGA-Depth
Builds on: Monodepth2 SurroundDepth MPViT
Closest papers by meaning
- Adaptive Fusion of Single-View and Multi-View Depth for Autonomous Driving (CVPR 2024 · similarity 0.86)
- RRT-MVS: Recurrent Regularization Transformer for Multi-View Stereo (AAAI 2025 · similarity 0.86)
- Synthetic-to-Real Self-supervised Robust Depth Estimation via Learning with Motion and Structure Priors (CVPR 2025 · similarity 0.85)
- Beware of Road Markings: A New Adversarial Patch Attack to Monocular Depth Estimation (NeurIPS 2024 · similarity 0.84)
- Tri-Perspective View Decomposition for Geometry-Aware Depth Completion (CVPR 2024 · similarity 0.84)
- DepthVanish: Optimizing Adversarial Interval Structures for Stereo-Depth-Invisible Patches (NeurIPS 2025 · similarity 0.84)
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.