Research map / 3D vision / Depth and stereo
Modeling Stereo-Confidence out of the End-to-End Stereo-Matching Network via Disparity Plane Sweep
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 a training-free stereo-confidence measure computed from a disparity volume obtained by plane-sweeping the right image through an external end-to-end stereo network, and shows it competes with learning-based confidence methods and improves them as an extra input modality.
Problem
Stereo confidence estimation for end-to-end stereo-matching networks whose cost volumes and internal features are inaccessible (e.g., safety-critical systems), where existing learning-based confidence methods depend on cost volumes and ground-truth-trained supervision.
Approach
Reinterpret the cost-profile-based confidence as a disparity-profile check: shifting the right image by k pixels should shift the predicted disparity by k, so the deviation of the zero-shift-anchored ideal ramp from the disparity volume measures matching ambiguity, computed purely from external disparity outputs.
Key results
- On K2012 with RAFT-Stereo, the proposed confidence (N=5, K=2) reaches AUC 0.4729, versus 0.7820 for LAF-Net* and 0.3862 for LAF-Net, while LAF-Net with the proposed cue as extra input (LAF-Net-dagger) reaches 0.3414.
- Using the proposed measure as an extra input to LAF-Net* (LAF-Net*-dagger) improves AUC on K2012 with RAFT from 0.7820 to 0.3563.
- On DrivingStereo with GANet, the proposed method achieves AUC 1.3013 versus 1.7419 for OTB and 1.4632 for OTB-online (self-supervised baselines), though supervised ConfNet-dagger reaches 0.9581.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Ours as an additional input modality to LAF-Net | LAF-Net with RAFT on K2012 · K2012 | AUC: 0.3862 -> 0.3414 | Helped |
| Add Ours as an additional input modality to LAF-Net* | LAF-Net* with RAFT on K2012 · K2012 | AUC: 0.7820 -> 0.3563 | Helped |
| Add Ours as an additional input modality to CCNN | CCNN with RAFT on K2012 · K2012 | AUC: 0.6997 -> 0.4129 | Helped |
| Increase disparity plane shifts from N=2 to N=7 (ablation) | Ours with RAFT on K2012 · K2012 | AUC: 0.7917 -> 0.4383 | Helped |
| Increase disparity plane shifts from N=2 to N=7 (ablation) | Ours with RAFT on K2012 · K2012 | Latency (1 GPU): 0.6490 -> 2.0735 | Hurt |
| Increase shifting step size from K=1 to K=8 (ablation) | Ours with N=3 and RAFT · K2012 | AUC: 0.5455 -> 0.6984 | Hurt |
| Use Ours instead of ConfNet | Confidence estimation with GANet · DrivingStereo | AUC: 1.7420 -> 1.3013 | Helped |
Surprising findings
- The training-free proposed measure with only disparity inputs outperforms several learning-based confidence methods on multiple datasets, including LAF-Net* that uses image and disparity inputs.
- LAF-Net* augmented with the proposed cue often beats full LAF-Net that uses the cost volume, suggesting the external cue can substitute for cost volume inputs.
Limitations
- Latency grows with the number of disparity plane shifts N, which is slow on a single GPU.
- Learning-based baselines and the learned variants require ground-truth-trained networks, and the proposed method's gains are not uniform across all stereo networks and conditions (some cases lose to LAF-Net variants).
Experimental setup
Datasets: KITTI 2012 KITTI 2015 Virtual KITTI 2 Middlebury 2014 DrivingStereo
Benchmarks: KITTI 2012 KITTI 2015 Virtual KITTI 2 Scene06 Middlebury 2014 DrivingStereo
Metrics: AUC of ROC (x100) latency (s)
Baselines: CCNN LFN ConfNet LGC-Net LAF-Net LAF-Net* OTB OTB-online
Builds on: LAF-Net CCNN ConfNet LFN LGC-Net
Closest papers by meaning
- Stereo Risk: A Continuous Modeling Approach to Stereo Matching (ICML 2024 · similarity 0.89)
- Any-Stereo: Arbitrary Scale Disparity Estimation for Iterative Stereo Matching (AAAI 2024 · similarity 0.87)
- Adaptive Multi-Modal Cross-Entropy Loss for Stereo Matching (CVPR 2024 · similarity 0.87)
- Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.87)
- Robust Synthetic-to-Real Transfer for Stereo Matching (CVPR 2024 · similarity 0.87)
- IINet: Implicit Intra-inter Information Fusion for Real-Time Stereo Matching (AAAI 2024 · 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.