Learning Category Aware Aggregation for Large Scale Point Cloud Semantic Segmentation

Abstract

Previous Point Cloud semantic segmentation methods lack the ability to fully capture the feature representation of the overlapping areas of different classes. Some works have identified the edge detection problem but they ignored the inner points of the objects in the scene. This research has applied an aggre gation strategy which processed features from same class and different classes equally. Each neighbor of the feature is identified first, if it belongs to the same class as the feature or a different class from the feature. These features are then dealt by two specific modules. We applied this strategy on semantic segmentation network to improve segmentation results. The results section proves the effectiveness of this methodology.