Vanilla ViT for Automotive Point Cloud Semantic Segmentation

Researchers introduced VaViT, a vanilla Vision Transformer for automotive LiDAR point cloud semantic segmentation.

With a redesigned tokenizer, lightweight decoder, and targeted augmentations, it closes much of the gap with U-Net-style models.

VaViT delivers strong results on nuScenes, SemanticKITTI, and Waymo Open Dataset.

The work suggests plain ViTs can compete with specialized 3D segmentation architectures, simplifying autonomous driving perception stacks.