EfficientLPS: Efficient LiDAR Panoptic Segmentation

TL;DR

EfficientLPS is a novel top-down architecture for LiDAR panoptic segmentation that effectively handles data challenges and achieves state-of-the-art results on nuScenes and SemanticKITTI datasets.

Contribution

The paper introduces a new EfficientLPS architecture with a shared backbone, scale-invariant heads, and a panoptic fusion module, improving LiDAR segmentation performance.

Findings

Sets new state-of-the-art on nuScenes and SemanticKITTI datasets.

Introduces a regularized pseudo labeling framework for unlabelled data.

Proposes a novel shared backbone with enhanced geometric modeling.

Abstract

Panoptic segmentation of point clouds is a crucial task that enables autonomous vehicles to comprehend their vicinity using their highly accurate and reliable LiDAR sensors. Existing top-down approaches tackle this problem by either combining independent task-specific networks or translating methods from the image domain ignoring the intricacies of LiDAR data and thus often resulting in sub-optimal performance. In this paper, we present the novel top-down Efficient LiDAR Panoptic Segmentation (EfficientLPS) architecture that addresses multiple challenges in segmenting LiDAR point clouds including distance-dependent sparsity, severe occlusions, large scale-variations, and re-projection errors. EfficientLPS comprises of a novel shared backbone that encodes with strengthened geometric transformation modeling capacity and aggregates semantically rich range-aware multi-scale features. It incorporates new scale-invariant semantic and instance segmentation heads along with the panoptic fusion module which is supervised by our proposed panoptic periphery loss function. Additionally, we formulate a regularized pseudo labeling framework to further improve the performance of EfficientLPS by training on unlabelled data. We benchmark our proposed model on two large-scale LiDAR datasets: nuScenes, for which we also provide ground truth annotations, and SemanticKITTI. Notably, EfficientLPS sets the new state-of-the-art on both these datasets.