Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Lite Any Stereo: Efficient Zero-Shot Stereo Matching
Junpeng Jing, Weixun Luo, Ye Mao, Krystian Mikolajczyk
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use hybrid 3D-then-2D cost aggregation instead of 2D only (ablation) | Lite Any Stereo · ETH3D (1.4M synthetic subset) | Bad 1.0: 6.48 -> 5.39 (2D vs 3D-2D); KITTI 2012 5.02 -> 4.78 | Mixed |
| place the 2D block before the 3D block (ablation) | Lite Any Stereo cost aggregation · ETH3D (1.4M synthetic subset) | Bad 1.0: 5.39 -> 28.85 | Hurt |
| enlarge the 3D kernel along the disparity dimension (ablation) | Lite Any Stereo cost aggregation · KITTI 2012 (1.4M synthetic subset) | D1: 4.38 -> 4.86 (kernel (11,3,3)) | Hurt |
| increase 3D block proportion beyond 4.8% (ablation) | Lite Any Stereo cost aggregation · Middlebury (1.4M synthetic subset) | Bad 2.0: 9.50 -> 10.34 (15.6%) | Hurt |
| use ConvNeXt layers instead of MobileNetV2 in the 2D block (ablation) | Lite Any Stereo · KITTI 2012 (1.4M synthetic subset) | D1: 4.78 -> 4.38; MACs 33.9 -> 32.8 | Improved |
| add self-distillation stage rather than direct data augmentation (ablation) | Lite Any Stereo · Middlebury (1.4M synthetic subset) | Bad 2.0: 11.14 (data aug.) -> 8.86 (know. dis.) | Improved |
| update teacher via EMA or hard copy instead of fixed teacher (ablation) | Stage 2 self-distillation · KITTI 2012 (1.4M synthetic subset) | D1: 3.64 -> 3.97 (EMA), 4.22 (hard copy) | Hurt |
| add stage 3 real-world pseudo-label distillation after stage 2 (ablation) | Lite Any Stereo · ETH3D (full training set) | Bad 1.0: 4.69 -> 3.53 | Improved |
| add stage 3 on real-world data (ablation) | Lite Any Stereo · Middlebury (full training set) | Bad 2.0: 7.03 -> 7.51 | Hurt |
| apply the three-stage training strategy to other efficient architectures | LightStereo-M · KITTI 2012 | D1: 4.34 -> 3.35 | Improved |
| apply the three-stage training strategy to BANet-2D | BANet-2D · Middlebury | Bad 2.0: 10.54 -> 10.30 (stage 3); KITTI 2012 4.34 -> 3.28 | Mixed |
| add low-quality or domain-specific data (e.g. Stereo4D, HRWSI) in stage 3 | Lite Any Stereo · zero-shot benchmarks | zero-shot accuracy | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- FoundationStereo — FoundationStereo: Zero-Shot Stereo Matching (CVPR 2025) · better on 0 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Any-Stereo: Arbitrary Scale Disparity Estimation for Iterative Stereo Matching (AAAI 2024 · similarity 0.91)
- Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.90)
- FoundationStereo: Zero-Shot Stereo Matching (CVPR 2025 · similarity 0.90)
- DEFOM-Stereo: Depth Foundation Model Based Stereo Matching (CVPR 2025 · similarity 0.88)
- MonSter: Marry Monodepth to Stereo Unleashes Power (CVPR 2025 · similarity 0.87)
- 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.
