DeepTemporalSeg: Temporally Consistent Semantic Segmentation of 3D LiDAR Scans

This website presents our work on semantic segmentation of a 3D LiDAR scan. We propose a new architecture called DBLiDARNet. The architecture is based on dense blocks and to limit the number of learnable parameters we use depth separable convolution in the decoder. To make predictions temporally consistent we propose a Bayes filter approach. The filter recursively estimates the current semantic state of a point by using prediction from current and previous scans.