RetSeg3D: Retention-based 3D semantic segmentation for autonomous driving

Download PDF
Open Webpage

Gopi Krishna Erabati, Helder Araújo

Published: 01 Jan 2025, Last Modified: 16 Jul 2025Comput. Vis. Image Underst. 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•RetSeg3D maintains the receptive field of far away sparse points.•Employs explicit decay to introduce spatial distance related prior information.•Captures long range dependencies & considers the varying-density property of LiDAR.•Experiments on three large-scale datasets: SemanticKITTI, nuScenes and Waymo.•Improved performance and efficiently runs in real time at 52.1 FPS on RTX 4090 GPU.