Multi-View Amodal Instance Segmentation Based on 3d Representation

Amodal instance segmentation (AIS) aims to predict the entire object, including both its visible and occluded parts. AIS is not only important for addressing occlusion issues in numerous industrial applications, but also an important step to develop a intelligent machine vision system. In this paper, a multi-view amodal instance Segmentation method based on 3D representations is developed. Image features are lifted into a 3D representation, multi-view features from different viewpoints are aggregated in 3D volumes, and the amodal masks and visible masks are decoded from the aggregated features. As we know, we are the first to introduce 3D representations to the AIS task. The experiments show that the developed method can achieve state-of-the-art performance for amodal instance segmentation.