Efficient CU-Based ROI Detection in H.266/VVC Video via Graph Convolutional Network

A region of interest (ROI) refers to a specific area within an image that attracts visual attention or contains critical information. Identifying and focusing on ROIs typically improves computational efficiency by avoiding redundant computation in non-informative areas. Therefore, detecting the ROI within a video frame is crucial for many multimedia applications. However, most existing methods rely on pixel-domain information and convolutional neural networks for ROI detection. The potential of using graph convolutional networks (GCNs) to exploit compressed-domain information for ROI detection remains underexplored. Therefore, this paper proposes a novel coding unit (CU)-level ROI detection method that employs a GCN and compressed-domain information from H.266/versatile video coding (VVC) encoded video. A video frame is constructed from CUs, each treated as a node in the graph. These nodes are connected via edges to establish a graph representation of the frame. Each node is associated with features such as geometric attributes, spatial-temporal position, coding mode, quantization parameter, motion characteristics, and residual statistics. The experimental results demonstrate that the proposed method effectively and efficiently detects ROI CUs while significantly reducing computation time compared to previous works.