[Submitted on 29 May 2026 (v1), last revised 26 Oct 2026 (this version, v2)]
Abstract:Plain Transformers have become the de-facto architecture for processing text, audio, image, and video, offering a unified backbone for multimodal learning. However, state-of-the-art architectures for point cloud semantic segmentation remain dominated by U-Nets architectures where convolutions are interleaved with local or windowed attentions. In this work, we show how to effectively leverage vanilla, non-hierarchical ViTs for the segmentation of large-scale automotive lidar scenes. We bridge the performance gap thanks to a carefully designed tokenizer, a lightweight decoder segmentation head, and tailored data augmentations. Our approach, VaViT, for Vanilla ViT, matches or exceeds the performance of state-of-the-art methods while maintaining the simplicity of ViT architecture. We provide extensive evaluations on nuScenes, SemanticKITTI, and Waymo Open Dataset to validate the efficiency of our method. Code and pre-trained models are available at this https URL.
Submission history
From: Nermin Samet [view email]
[v1] Fri, 29 May 2026 11:47:02 UTC (34,927 KB)
[v2] Mon, 26 Oct 2026 13:56:00 UTC (35,887 KB)
