Segmentation and Classification of 3D Point Clouds Using Deep Pointwise Residual Networks

Abstract

Deep learning in computer vision domain plays an important role in the last few years. Its performance surpasses previous approaches based on handcrafted features. Deep learning based approaches yield better performance in scene understanding tasks. Scene understanding tasks have great importance in the robot navigation for the indoor and outdoor environment. Recently, scene understanding tasks captured the attention of the researchers towards 3D domain, because 3D data provides rich information of geometry about the objects and scenes. Deep learning due to its power of extracting discriminative hierarchical features from input became popular in the 3D do main. 3D point clouds are compact and computationally less expensive as compare to volumetric 3D data. 3D point clouds are memory efficient too. Recently, fully pointwise convolutional neural network has been designed for unstructured 3D point clouds [1]. The designed 3D kernel can be placed on each point in 3D point cloud. But training this network as deeper network was an open challenge because it caused performance degradation. But this degradation problem can be resolved with training it through residual learning, because residual learning allows the network to become deeper without sacrificing performance. Here in this paper we train deep pointwise CNN using residual learning and presented different experiments. Training of deep architecture with residual learning in the domain of unstructured 3D point clouds is still not performed before. Our qualitative results show that we are successfully able to overcome this problem of performance degradation due to network depth. We evaluated our deep network on scene segmentation and object recognition task. On scene segmentation task, for Stanford 3D indoor scene(S3DIS) dataset our deep network gives competitive performance as compare to many existing state-of-the-art architectures.