Research map / 3D vision / Depth and stereo
Any-Stereo: Arbitrary Scale Disparity Estimation for Iterative 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 Any-Stereo, an implicit-neural-representation disparity upsampling module that replaces convolutional upsampling in iterative stereo models (RAFT-Stereo, IGEV-Stereo), enabling arbitrary-scale disparity output.
Problem
Iterative stereo matching maintains regularized disparity at low resolution, and convolutional upsampling to full resolution fails to recover high-frequency details lost during downsampling, limiting full-resolution accuracy.
Approach
Models the neighbor-combination mask for disparity upsampling as a continuous implicit neural function (LIIF-style) over hidden-state latent codes, augmented with intra-scale similarity unfolding (cheap self-similarity features instead of costly feature unfolding) and cross-scale feature alignment from multi-scale image context via pixel-unshuffle downsampling blocks, plus multi-scale training wit
Key results
- On KITTI 2015 the Any-IGEV variant improves D1-fg (Noc) from 2.62% to 2.27% over IGEV-Stereo and ranks 1st on the leaderboard at submission time.
- Any-RAFT reduces RAFT-Stereo's parameters from 11.23M to 10.90M while improving SceneFlow >1px error from 7.87% to 6.80%.
- When inputs are downsampled to 25%, Any-RAFT yields about 50% lower error than RAFT-Stereo on KITTI15 and Middlebury (e.g. 10.6 vs 21.0 on KITTI15).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace IGEV-Stereo upsampling with AnyStereo | IGEV-Stereo · KITTI 2015 | D1-fg (Noc) (%): 2.62 -> 2.27 | Helped |
| Replace IGEV-Stereo upsampling with AnyStereo | IGEV-Stereo · KITTI 2012 | Out-All (%): 1.44 -> 1.41 | Helped |
| Replace convolutional decoding with INMF (ablation) | RAFT-Stereo upsampling · SceneFlow test | EPE (px): 0.88 -> 0.83 | Helped |
| Use multi-scale training rather than fixed-scale training (ablation) | INMF · SceneFlow test | EPE (px): 0.83 -> 0.83 | No clear effect |
| Add ISU (ablation) | INMF · SceneFlow test | EPE (px): 0.77 -> 0.75 | Helped |
| Set ISU window size P to 3 instead of 5 (ablation) | INMF with ISU · SceneFlow test | EPE (px): 0.75 -> 0.76 | Hurt |
| Add CFA (ablation) | INMF · SceneFlow test | EPE (px): 0.80 -> 0.75 | Helped |
| Add SCA to CFA (ablation) | INMF with CFA · SceneFlow test | EPE (px): 0.76 -> 0.75 | Helped |
Limitations
- Gains on full-resolution KITTI benchmarks are modest for the IGEV baseline in several metrics (e.g., Out-All 1.44 to 1.41).
- Multi-scale training is slightly inferior to fixed-scale training on SceneFlow at full resolution.
- Intra-scale similarity unfolding is a cheaper approximation of feature unfolding, not the full LIIF-style local ensemble.
Experimental setup
Datasets: SceneFlow KITTI 2012 KITTI 2015 Middlebury
Benchmarks: KITTI 2015 KITTI 2012 Middlebury SceneFlow
Metrics: D1-bg D1-fg D1-all Out-Noc Out-All EPE >1px Params (M) Runtime (s)
Baselines: RAFT-Stereo IGEV-Stereo DLNR PSMNet GANet AANet LEAStereo SMD-Net ACVNet LIIF Bicubic interpolation
Builds on: RAFT-Stereo IGEV-Stereo LIIF
Closest papers by meaning
- Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.88)
- Mono2Stereo: A Benchmark and Empirical Study for Stereo Conversion (CVPR 2025 · similarity 0.88)
- DEFOM-Stereo: Depth Foundation Model Based Stereo Matching (CVPR 2025 · similarity 0.87)
- Modeling Stereo-Confidence out of the End-to-End Stereo-Matching Network via Disparity Plane Sweep (AAAI 2024 · similarity 0.87)
- Selective-Stereo: Adaptive Frequency Information Selection for Stereo Matching (CVPR 2024 · similarity 0.87)
- MonSter: Marry Monodepth to Stereo Unleashes Power (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.