UPOFSE 2012 - Workshop on Unsolved Problems in Optical Flow and Stereo Estimation

More than 1500 articles on optical flow have been published since 1980 in four major journals alone. Only few assess the performance of existing methods. Main reasons are a general lack of widely accepted best practises on performance characterization and diverse ground truth data. Therefore, it also becomes difficult to measure scientific advances. This workshop aims at bringing together two groups of scientists: those who use stereo and optical flow methods in their scientific work as a tool and those who devise new models and optimization strategies. We would like to create a chance to discuss priorities and perspective between scientists in both of these groups.
Established benchmarks like the Middlebury datasets have had a tremendous impact on the research activity in dense correspondence estimation, but current benchmarks are nearly solved, and new challenges are needed. However, creating ground truth for real image sequences is difficult, and no established techniques exist for outdoor sequences. Synthetic sequences with ground truth bear promise, but existing such datasets have limited realism, and it is not known whether all aspects of real data can be adequately modeled so that performance results on synthetic data are meaningful.
Goals and topics
The goal of this workshop is to make progress on unsolved problems in optical flow and stereo estimation, even in the absence of test data with ground truth. We provide challenging real-world data, including stereo sequences taken from moving vehicles, on which existing methods tend to fail completely. Sample problems include large motions, non-Lambertian surfaces, extreme lighting variations, motion blur, etc. We invite contributions that address these challenges, as well as contributions in all related areas. Specific topics of interest include, but are not limited to, the following:
advances in stereo and flow methods that address unsolved real-world challenges
confidence estimation and model violation detection
novel performance evaluation methods, especially without ground truth
creation and validation of highly realistic synthetic test data
open source implementation, comparison and discussion of existing methods
The Robust Vision Challenge
Together with Robert Bosch GmbH we recorded real-world outdoor stereo sequences taken from moving vehicles during different seasons, on which both stereo and optical flow methods often fail due to model violations. Errors tend to be significant and noticeable, so that scientific advance in the modeling or detection can be observed even without ground truth information. We pose a challenge for each sequence, whose results will be evaluated by a Jury consisting of experts working for well-known companies.
Please visit the Robust Vision Challenge Website to learn more.
Invited Speakers
Daniel Cremers (TU München, Germany)