Research map / 3D vision / Depth and stereo
LoS: Local Structure-Guided Stereo Matching
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 LoS, an optimization-based stereo matcher that uses local structure information (slant-plane gradients, disparity offsets, neighbour relations) bootstrapped from monocular depth to guide propagation-based disparity updates inside a GRU framework.
Problem
Stereo matching in challenging areas (out-of-view left regions, occlusions, textureless areas, edges) where appearance alone is ambiguous, and GRU-based optimization methods need many iterations to converge.
Approach
Represents local structure as a unified LSI combining slant-plane disparity gradients, per-neighbour disparity offsets and learned local relations, and uses it in a parameter-free Local Structure-Guided Propagation (LSGP) that propagates low-uncertainty disparities to high-uncertainty pixels; the LSI is initialised from MiDaS monocular depth and refined iteratively by a multi-level ConvGRU.
Key results
- LoS achieves the best Bad1.0 on ETH3D (1.03 all) and the lowest AvgErr-all on Middlebury (1.75), among published methods.
- On UnrealStereo4K, LoS reaches overall Bad2.0 15.0 and AvgErr 6.28 versus RAFT-Stereo's 15.7 and 14.11, and LSGP reduces AvgErr from 18.81 to 6.28 relative to the variant without LSGP.
- LoS runs at 0.31s at 960x640 on RTX4090 versus 0.43s for RAFT-Stereo, 0.41s for CREStereo and 0.47s for IGEV.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LoS instead of CREStereo | stereo matching · ETH3D | Bad1.0 all: 1.09 -> 1.03 | Helped |
| use LoS instead of DLNR | stereo matching · Middlebury | Bad2.0 all: 6.98 -> 8.03 | Hurt |
| use LoS instead of DLNR | stereo matching · Middlebury | AvgErr all: 1.91 -> 1.75 | Helped |
| use LoS instead of IGEV | stereo matching · KITTI 2015 | D1-all all: 1.59 -> 1.65 | Hurt |
| use LoS instead of IGEV | stereo matching · KITTI 2012 | out-3 all: 5.00 -> 4.45 | Helped |
| add LSGP (ablation) | LoS · UnrealStereo4K overall | AvgErr: 18.81 -> 6.28 | Helped |
| add LSGP (ablation) | LoS · UnrealStereo4K overall | Bad2.0: 19.9 -> 15.0 | Helped |
Surprising findings
- Local relations learned by the network tend to point pixels on the left side of the image (class 1 regions) toward right-side neighbours, showing the model learns geometry-consistent neighbour selection.
Limitations
- LSI struggles with extremely complex structures such as inter-occlusions and dense vegetation in KITTI.
- No constraints on disparity offset residuals, causing residual self-offsets that limit adjusting the number of LSGP iterations at inference.
- High-uncertainty pixels may still affect low-uncertainty pixels, slightly hurting Bad-x.0 metrics.
Experimental setup
Datasets: SceneFlow CREStereo dataset MPI-Sintel FallingThings Instereo2K ETH3D training set Middlebury training set KITTI 2012 KITTI 2015
Benchmarks: ETH3D Middlebury KITTI 2015 KITTI 2012 Robust Vision Challenge UnrealStereo4K
Metrics: Bad1.0 AvgErr Bad2.0 D1-all out-3 runtime
Baselines: RAFT-Stereo CREStereo IGEV DLNR CroCo-Stereo HITNet ACVNet LEAStereo CREStereo++ RVC CFNet RVC
Builds on: RAFT-Stereo CREStereo HITNet IGEV MiDaS
Closest papers by meaning
- DEFOM-Stereo: Depth Foundation Model Based Stereo Matching (CVPR 2025 · similarity 0.86)
- Dual-Level Precision Edges Guided Multi-View Stereo with Accurate Planarization (AAAI 2025 · similarity 0.86)
- Any-Stereo: Arbitrary Scale Disparity Estimation for Iterative Stereo Matching (AAAI 2024 · similarity 0.86)
- MonSter: Marry Monodepth to Stereo Unleashes Power (CVPR 2025 · similarity 0.86)
- Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.86)
- Stereo Risk: A Continuous Modeling Approach to Stereo Matching (ICML 2024 · similarity 0.86)
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.