NeurIPS 2026 Competition – RoCo-Spring: The Robust Correspondence Challenge

August 11, 2026; Christina Warren

At this year's Annual Conference on Neural Information Processing Systems (NeurIPS 2026), held December 6–12 in Sydney, Australia, researchers from Projects B04 and A07 – together with colleagues from Mannheim, Saarbrücken, Beijing and NVIDIA – will organize the RoCo-Spring Competition – a challenge for jointly evaluating the accuracy and robustness of methods for optical flow, scene flow and stereo estimation. With only 16 out of 89 proposals accepted (18%), the NeurIPS competition track was particularly competitive this time.

The RoCo-Spring Competition is based on Spring and Robust Spring (https://spring-benchmark.org/) – two high detail high resolution benchmarks that have been published by SFB-TRR 161 researchers at CVPR 2023 in Vancouver, Canada and ICLR 2026 in Rio de Janeiro, Brazil, respectively. The challenge comprises four tracks: three tracks aiming at developing novel, more robust approaches for computing optical flow, scene flow and stereo, and a fourth track that aims at exploring the robustness of computer vision methods under distribution shifts in general. The competition welcomes the submissions of papers in all four tracks, which will be peer reviewed and in case of acceptance presented at the corresponding NeurIPS workshop. For the workshop, Jia Deng (Princeton University) and Fatih Porikli (Qualcomm AI Research) have been confirmed as keynote speakers.

The RoCo-Spring Competition is supported by the SFB-TRR 161 "Quantiative Methods for Visual Computing".