Title: A Robust Approach to CVPR 2026’s 8th UG2+ Challenge Track 3: Dynamic Object Segmentation in Turbulence

Abstract:

This paper outlines our winning strategy for Track 3 of the 8th UG2+ Challenge at CVPR 2026, which focuses on Dynamic Object Segmentation in Turbulence (DOST). Our methodology leverages the robust baseline framework known as Segment Any Motion (SegAnyMo), renowned for its advanced capabilities in mask generation and motion tracking. To enhance segmentation accuracy amidst severe atmospheric distortions, we implemented two primary advancements.

First, we adopted a data-centric domain adaptation approach. We augmented our training corpus by integrating specific sequences from the DAVIS dataset with a portion of the DOST dataset. Furthermore, we applied simulated atmospheric fluctuation degradations to this expanded data, thereby strengthening the model’s resilience to complex geometric distortions.

Second, we developed a spatio-temporal post-processing module. This refinement stage successfully eliminates persistent false foregrounds connected at the boundaries and transient fragmented noise, all while ensuring that authentic small targets are retained and original individual labels remain consistent across frames. Through the integration of these strategies, our method achieved second place in the competition.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC