A New Semantic Edge Aware Network for Object Affordance Detection

This paper presents a new object affordance detection framework for robotic applications. In robotic manipulation tasks, the robot needs the accurate edge information of objects. Therefore, in order to improve the edge quality of the affordance detection results, semantic edge detection is introduced into our affordance detection framework. In order to take full advantage of the duality between affordance detection task and semantic edge detection task, our framework employs two key components to couple these two tasks: spatial gradient fusion module and shared gradient attention module. In the spatial gradient fusion module, the gradient features derived from the segmentation block are fused with the features output by the edge block to suppress non-semantic edges in semantic edge detection. In the shared gradient attention module, the edge consistency between affordance detection results and semantic edge detection results is enhanced by sharing gradient attention weights. Our experiments show that our framework can output affordance detection results with better edge quality. In particular, our method achieves state-of-the-art performance on the IIT-AFF benchmark, in terms of both affordance mask (F-score) and semantic edge (F-score) quality, improving by 0.52% and 3.22% over strong baselines. Furthermore, we demonstrate the effectiveness of our framework in real robotic applications.