Research map / 3D vision / Depth and stereo
A Global Depth-Range-Free Multi-View Stereo Transformer Network with Pose Embedding
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 GD-PoseMVS, a depth-range-free multi-view stereo transformer that jointly processes all source views via multi-view disparity attention with 3D pose embedding and uncertainty-driven disparity hidden states.
Problem
Learning-based multi-view stereo is sensitive to the depth range prior, and existing depth-range-free methods work pair-wise, ignoring inter-image correspondences among source views.
Approach
Estimates epipolar disparity flows for all source views simultaneously using a Multi-view Disparity Attention module (intra-image self-attention plus inter-view cross-attention) with 3D pose embedding encoding relative pose and pixel geometry, plus per-source hidden states updated by a variance-based uncertainty module; initialization picks the midpoint of the valid epipolar search range without d
Key results
- On DTU evaluation set, the method reaches 0.335 mm overall, the best among depth-range-free methods (DispMVS 0.339, RAMDepth 0.362).
- On Tanks and Temples, mean F-score of 35.95 on advanced and 59.27 on intermediate, best among scale-agnostic methods (DispMVS 34.90 / 59.07).
- Under a wide depth range (141.6mm to 2805mm) the method keeps 0.335 mm overall while DispMVS degrades to 0.493 and CER-MVS to 12.073.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add 3D pose embedding (ablation) | Multi-view attention for MVS · DTU | Overall (mm): 9.46% | Helped |
| Add uncertainty estimation and disparity hidden state updating (ablation) | Disparity feature encoding for MVS · DTU | Overall (mm): 9.95% | Helped |
| Add uncertainty estimation (ablation) | Disparity feature encoding for MVS · DTU | Overall (mm): 6.16% | Helped |
| Add disparity hidden state updating (ablation) | Disparity feature encoding for MVS · DTU | Overall (mm): 7.46% | Helped |
| Use the proposed method without depth-range priors | Ours · DTU with varied depth ranges | Overall (mm): 0.335 -> 0.335 | Helped |
| Replace ground-truth depth range with COLMAP-estimated range | Ours · DTU | Overall (mm): 0.335 -> 0.335 | No clear effect |
Surprising findings
- Depth-range-based SOTA methods (GeoMVS, MVSFormer++) have lower DTU overall error than the proposed method when given accurate depth ranges, indicating a remaining gap.
Limitations
- Cannot run in real time (30 FPS), limiting mobile or time-sensitive use.
- Performance gap versus SOTA depth-range-based cost-volume methods on DTU and Tanks and Temples.
Experimental setup
Datasets: DTU BlendedMVS Tanks and Temples
Benchmarks: DTU Tanks and Temples
Metrics: Accuracy (mm) Completeness (mm) Overall (mm) F-score
Baselines: DispMVS RAMDepth IterMVS CER-MVS GeoMVS MVSFormer++ MVSTER GBi-Net TransMVSNet UniMVSNet COLMAP Gipuma
Builds on: DispMVS IterMVS VideoFlow MVSNet
Closest papers by meaning
- MVS2D: Efficient Multi-View Stereo via Attention-Driven 2D Convolutions (CVPR 2022 · similarity 0.90)
- Rethinking Disparity: A Depth Range Free Multi-View Stereo Based on Disparity (AAAI 2023 · similarity 0.90)
- MVSFormer++: Revealing the Devil in Transformer's Details for Multi-View Stereo (ICLR 2024 · similarity 0.89)
- RIAV-MVS: Recurrent-Indexing an Asymmetric Volume for Multi-View Stereo (CVPR 2023 · similarity 0.89)
- RRT-MVS: Recurrent Regularization Transformer for Multi-View Stereo (AAAI 2025 · similarity 0.88)
- MVSAnywhere: Zero-Shot Multi-View Stereo (CVPR 2025 · 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.