Priors for Stereo Vision under Adverse Weather Conditions

Gehrig, Stefan; Reznitskii, Maxim; Schneider, Nicolai; Franke, Uwe; Weickert, Joachim · 2013 · OpenAlex-citations

DOI: 10.1109/iccvw.2013.39

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of robust stereo vision in autonomous driving scenarios under adverse weather conditions, such as rain, snow, and night. While Semi-Global Matching (SGM) is a standard method for real-time 3D reconstruction, it suffers from temporal noise, disparity outliers, and false positives in these difficult environments. The authors identify two specific issues: the inability to recover from transient occlusions (like windshield wipers) and a disparity distribution that generates false positives in the critical near-field region. To solve this, the paper introduces two new priors—a temporal prior and a scene prior—and integrates them into both Graph Cut (GC) and SGM frameworks. The methodology begins by reformulating SGM into a probabilistic Bayesian framework to facilitate the integration of priors. The temporal prior leverages known ego-motion and a static-world assumption to predict disparity maps from previous frames. To handle moving objects and ego-motion errors, the authors compute a confidence map using sparse optical flow residuals and disparity gradients, ensuring that the prior is only applied where the static-world assumption holds. The scene prior addresses the hyperbolic nature of disparity-distance relations, which causes outliers to cluster in nearby 3D space. It imposes a prior that expects a uniform distribution of points in 3D space, effectively penalizing false positives in the immediate vicinity of the vehicle. These priors are implemented as additional terms in the energy minimization function for GC and as modifications to the data cost volume for SGM. Experiments were conducted on the KITTI dataset and a proprietary 3,000-frame highway database containing adverse weather conditions with sparse ground truth labels (stixels). On the KITTI dataset, the priors maintained performance comparable to standard SGM without degrading accuracy in good weather. On the adverse weather database, the results demonstrated significant improvements. The combination of both priors (TempScenePrior) reduced the false positive point rate by a factor of five compared to the SGM baseline and achieved the lowest false positive stixel counts. Furthermore, the temporal prior significantly improved detection rates, particularly in frames with windshield wiper passes, by inpainting missing data from previous frames. The authors also outperformed the winner of the ECCV Robust Vision Challenge (iSGM) on this specific database. The significance of this work lies in its ability to stabilize stereo vision under conditions where image data alone is insufficient. By incorporating temporal consistency and 3D scene expectations, the method reduces false positives in the critical driving corridor while maintaining high detection rates. The approach is computationally efficient, adding only 30ms of processing time, making it suitable for real-time driver assistance systems. The paper demonstrates that transferring these priors back to the efficient SGM algorithm allows for robust performance without sacrificing the speed required for production vehicles.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• weather rain fog snow