Based on FFB6D model, we constrained ROI of RGBD images with YOLOv5 detection model. Also add skip connection from encoder to decoder of FFB6D model. For multi-level fusion with skip connection, we use channel wise attention to filter meaningful feature.