Detection-Tracking Guided Video Instance Matting

Video matting and video instance segmentation cannot distinguish alpha mattes in a scenario featuring multiple instances. We propose a video instance matting method guided by detection and tracking combined with YOLOv8 and DeepSort.Then, we incorporate the Convolutional Block Attention Module (CBAM) to enhance attention distribution in the channel and spatial dimensions. It enables accurate tracking of multiple instances, then integrated with a GhostNet-based semantic matting model to obtain matting results for multiple instances in the video. Experiments conducted on the VIM-NEW dataset demonstrate that the proposed algorithm effectively addresses occlusion issues among multiple targets, achieves high matting accuracy, and reduces the model's parameter size.