3D Semantic Segmentation for Large-Scale Scene Understanding

Kiran Akadas, Shankar Gangisetty; Proceedings of the Asian Conference on Computer Vision (ACCV) Workshops, 2020

3D semantic segmentation is one of the most challenging events in the robotic vision tasks for detection and identification of var- ious objects in a scene. In this paper, we solve the task of semantic segmentation to classify and assign every point in the scene with an as- sociated label. We propose a lightweight semantic segmentation network for large-scale point clouds which consists of grid subsampling, dilated convolutions, and Gaussian error linear unit activation for gaining better performance. The dilated convolutions increase the receptive field while reducing the number of parameters, making proposed network faster and computationally more efficient with reduced number of parameters. Ad- ditionally, we use conditional random field as post processing method to boost the performance of proposed semantic segmentation network. We perform an exhaustive quantitative analysis of the proposed network on SOTA datasets, namely, SHREC 2020 street scenes dataset [1], S3DIS [2] and SemanticKITTI [3]. We show that proposed semantic segmentation network performs effectively and efficiently compared to SOTA methods.