Research map / Label-efficient and robust learning / Self-supervised and contrastive
DualNet: Robust Self-Supervised Stereo Matching with Pseudo-Label Supervision
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 DualNet, a two-step self-supervised stereo matching framework that trains a CFNet teacher with feature-metric and data-augmentation consistency losses, then trains a student on geometrically filtered pseudo disparity probability distributions.
Problem
Self-supervised stereo matching relies on the photo-metric consistency assumption, which is violated by low texture, occlusion, reflections, and illumination changes, yielding ambiguous supervision and inferior accuracy relative to supervised methods.
Approach
Replaces reliance on photo-metric loss with feature-metric and data-augmentation consistency losses for a teacher, then distills a student via uni-modal pseudo disparity probability distributions built from left-right-check-filtered teacher outputs, iterated for multiple rounds.
Key results
- DualNet (Step 2) reaches Out-Noc 2.46% and Out-All 5.25% on KITTI 2015 ... with KITTI 2012 Out-All 2.59%, ranking 1st among self-supervised methods in Table 1.
- On ETH3D, DualNet Step 2 achieves Bad 1.0 of 3.05%, outperforming supervised CFNet (3.31%) and GANet-RSSM (3.27%) in Table 2.
- Teacher (step 1) reaches KITTI 2015 Bad 3.0 of 3.31%, and two rounds of pseudo-label supervision lower it to 2.32%/2.36% range in Table 5 ablation (round 2 KITTI 2015 Bad 3.0 2.36).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add feature-metric consistency loss (ablation) | DualNet self-supervised teacher training · KITTI 2015 | Bad 3.0: 4.10 -> 3.88 | Helped |
| add data augmentation consistency loss (ablation) | DualNet self-supervised teacher training · KITTI 2015 | Bad 3.0: 3.88 -> 3.31 | Helped |
| add feature-metric consistency loss (ablation) | DualNet self-supervised teacher training · Middlebury | Bad 2.0: 12.8 -> 11.0 | Helped |
| add data augmentation consistency loss (ablation) | DualNet self-supervised teacher training · Middlebury | Bad 2.0: 11.0 -> 9.25 | Helped |
| add feature-metric consistency loss (ablation) | DualNet self-supervised teacher training · ETH3D | Bad 1.0: 8.13 -> 6.32 | Helped |
| add data augmentation consistency loss (ablation) | DualNet self-supervised teacher training · ETH3D | Bad 1.0: 6.32 -> 3.47 | Helped |
| add a second pseudo-label supervision round (ablation) | DualNet · KITTI 2015 | Bad 3.0: 2.68 -> 2.36 | Helped |
| add a third pseudo-label supervision round (ablation) | DualNet · KITTI 2015 | Bad 3.0: 2.36 -> 2.32 | Helped |
| add a second pseudo-label supervision round (ablation) | DualNet · Middlebury | Bad 2.0: 8.00 -> 7.55 | Helped |
| add a third pseudo-label supervision round (ablation) | DualNet · Middlebury | Bad 2.0: 7.55 -> 7.64 | Hurt |
| add a second pseudo-label supervision round (ablation) | DualNet · ETH3D | Bad 1.0: 2.10 -> 1.67 | Helped |
| add a third pseudo-label supervision round (ablation) | DualNet · ETH3D | Bad 1.0: 1.67 -> 1.61 | Helped |
Surprising findings
- The student model surpasses its teacher by a large margin and even outperforms several supervised methods despite using no ground-truth labels.
- Iterating pseudo-label supervision improves quality of pseudo labels, and performance saturates after about two rounds.
Limitations
- Relies on a teacher whose pseudo labels still contain outliers; filtering threshold chosen heuristically.
- Middlebury test benchmark allows only one entry per paper, so results are reported on training public tables instead.
Experimental setup
Datasets: KITTI 2012 KITTI 2015 Middlebury ETH3D SceneFlow
Benchmarks: KITTI 2012 KITTI 2015 Middlebury ETH3D
Metrics: Out-Noc Out-All Avg-Noc Avg-All D1-bg D1-fg D1-All Bad 1.0 Bad 2.0 Bad 3.0
Baselines: SssMnet Segstereo Flow2Stereo OASM PASMNet PAM F2S ChiT CFNet GANet-RSSM UPFNet DispNet
Builds on: CFNet photo-metric consistency loss KL divergence distillation contrastive learning with hard positives
Closest papers by meaning
- Consistency-aware Self-Training for Iterative-based Stereo Matching (CVPR 2025 · similarity 0.90)
- Self-Supervised Learning of Depth Inference for Multi-View Stereo (CVPR 2021 · similarity 0.89)
- Flow2Stereo: Effective Self-Supervised Learning of Optical Flow and Stereo Matching (CVPR 2020 · similarity 0.89)
- Local Similarity Pattern and Cost Self-Reassembling for Deep Stereo Matching Networks (AAAI 2022 · similarity 0.87)
- Robust Synthetic-to-Real Transfer for Stereo Matching (CVPR 2024 · similarity 0.86)
- Hierarchical Object-Aware Dual-Level Contrastive Learning for Domain Generalized Stereo Matching (NeurIPS 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.