De-Supervision in Camouflaged Videos

Luca Alessandrini, Antonino Maria Rizzo, Luca Magri, Giacomo Boracchi, Federica Arrigoni; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026, pp. 8401-8410

The aim of Zero-shot Video Camouflaged Object Segmentation (ZVCOS) is to automatically separate the foreground subject from the background, in videos where the subject is camouflaged within the environment, without any user intervention. Camouflaged videos represent the most challenging setting for video object segmentation, due to the minimal appearance-based cues available for the camouflaged subject, which closely resembles its surroundings. Consequently, ZVCOS has received limited research attention, primarily due to the scarcity of annotated datasets, with most existing approaches focusing on the supervised scenario. In this paper we introduce a simple but effective framework, named DeSC-V, that operates in an unsupervised manner. On one side, we exploit prior knowledge on the camouflaged subjects' appearance, roughly estimated from an image segmentation network. On the other side, we enhance such prior knowledge by taking advantage of the temporal information coming from close/distant time frames through the Optical Flow, which enforces global coherence among the estimated masks within the video: this allows us to address the challenge of transferring information from images to a video in a principled way. Experimental results on camouflaged datasets show that DeSC-V is effective, outperforming its closest competitor.