Object-Contextual Representations for PointNet

The classical network PointNet has achieved good results in classification and part segmentation for point clouds, but poor semantic segmentation results it gets with ignoring the context information of the point clouds in large scenes. In this paper, we present a simple approach, called object-contextual representations, characterizing a point by exploiting the representation of the corresponding object class. Firstly, object regions under the supervision of the ground truth are got by learning. Secondly, we compute the object region representation by aggregating the representations of the points lying in the object region. Last, the relation between each point and each object region are computed, and the representation of each point is augmented with the object-contextual representation, which is a weighted aggregation of all the object region representations. Furthermore, we adopt APNB module which leverages a pyramid sampling module into the non-local block to largely reduce the computation and memory consumption without sacrificing the performance. We believe that this method has strong versatility, and can also improve the performance with other popular networks.