Fully Convolutional Object Depth Prediction for 3D Segmentation from 2.5D Input

抄録

3D object segmentation is a crucial ability of machine to percept the real world environment, and previous works on this problem used 2D segmentation using rgb-d sensors. In environments with heavy occlusions, however, there are fragments in segmentation results even with mapping in multiple views. We tackle this problem with object depth prediction by convolutional networks. In our method, the occluded surface depth of objects is predicted from input rgb images, and 3d points are generated from prediction and input depth. We use datasets with 3D annotations for training, and show the performance and real-time efficiency our method.