YUV20K: a complexity-driven benchmark and trajectory-aware alignment model for video camouflaged object detection

Abstract The development of video camouflaged object detection (VCOD) is currently constrained by the scarcity of challenging benchmarks and the lack of robustness of models in the face of erratic motion dynamics. Existing methods frequently encounter challenges related to motion-induced appearance instability and temporal feature misalignment, particularly in complex motion scenarios. To address the issue of data bottleneck, we present YUV20K, a pixel-level annotated, complexity-driven VCOD benchmark. It is composed of 24,295 annotated frames, which are organized into 91 scenes and 47 species. It focuses on challenging scenarios, including large displacement motion, camera motion, and four additional complex-motion scenarios. In terms of methodology, a novel framework is proposed, comprising two key modules: motion feature stabilization (MFS) and trajectory-aware alignment (TAA). The MFS module utilizes frame-agnostic semantic basis primitives to stabilize features, while the TAA module leverages trajectory-guided deformable sampling to ensure precise temporal alignment. Extensive experiments demonstrate that our method achieves strong performance on existing datasets and establishes a competitive baseline on the challenging YUV20K.