Semantic Encoding Supervises LiDAR Semantic Segmentation

Semantic segmentation using LiDAR is a fundamental aspect in perception for autonomous driving. Conventional training methodologies commonly employ one-hot encoding for labels, followed by a softmax activation and a cross-entropy loss function to facilitate model training. Nevertheless, this approach exhibits several limitations: 1) information loss due to the use of one-hot encoding, 2) overconfidence issues induced by the softmax function, and 3) inability to accurately predict unknown or unlabeled classes. In this paper, we propose to use a pre-trained semantic sentence encoding model for label encoding to address these shortages. We theoretically demonstrate that the proposed method maximizes the accuracy, mitigates the overconfidence, and enhances the mean intersection over union (mIoU) metric. Empirical experiments further validate our approach, showing that it not only improves the perception performance but also enables an effective zero-shot learning and a cross-dataset transfer with minimal trade-offs.