SKIRP: an efficient inference strategy for video camouflaged object detection

Camouflaged object detection (COD) in video sequences poses significant challenges due to the visual ambiguity between objects and backgrounds. Although existing approaches often rely on large-scale backbone networks for fine-grained detection, their dense full-frame inference strategy leads to considerable computational overhead. To address this, we introduce SKIRP (Sparse Keyframe Inference and Region-of-Interest Propagation), an efficient inference mechanism tailored for video-based COD tasks. SKIRP performs full-frame inference on a sparse set of keyframes to identify high-quality regions of interest (ROIs), which are then propagated to intermediate non-keyframes via interpolation. On non-keyframes, inference is conducted solely within the propagated ROIs with a significantly reduced input size and computation. SKIRP enables efficient localized segmentation without dense per-frame processing by combining sparse inference and region-level propagation. The experimental results demonstrate around a 30% reduction in the computation compared to the base detector while maintaining a comparable segmentation quality. This design balances accuracy and efficiency, making it well-suited for scalable COD in long videos and edge-deployable scenarios.