Abstract:
3D point cloud segmentation provides a high-level semantic understanding of object structure that is valuable in applications such as medicine, robotics and self-driving. In this paper, we propose an Adversarial Graph Convolutional Network for 3D point cloud segmentation. Many current networks encounter problems such as low segmentation accuracy and high complexities due to their crude network architectures and local feature aggregation methods. To overcome these problems, we propose a) a graph convolutional network (GCN) in an adversarial learning scheme where a discriminator network provides a segmentation network with informative information to improve segmentation accuracy and b) a graph convolution, GeoEdgeConv, as a means of local feature aggregation to improve segmentation accuracy and space and time complexities. By using an embedding L2 loss as an adversarial loss, the proposed network is learned to reduce noisy labels by enforcing the consistency between neighbouring labels. Preserving geometric structures over convolution layers by using both point and relative position features, GeoEdgeConv helps learn fine details of complex structures, and thus improves segmentation accuracy in boundaries and reduces label noise inside a class without increased computational complexity. Experiments on ShapeNet Part demonstrate that our model outperforms the state-of-the-art (SOTA) with lower complexity and it has strong prospects in applications requiring low power but high segmentation performance.
