LiDAR-based Panoptic Segmentation via Dynamic Shifting Network

Abstract

With the rapid advances of autonomous driving, it becomes critical to equip its sensing system with more holistic 3D perception. However, existing works focus on parsing either the objects (e.g. cars and pedestrians) or scenes (e.g. trees and buildings) from the LiDAR sensor. In this work, we address the task of LiDAR-based panoptic segmentation, which aims to parse both objects and scenes in a unified manner. As one of the first endeavors towards this new challenging task, we propose the Dynamic Shifting Network (DS-Net), which serves as an effective panoptic segmentation framework in the point cloud realm. In particular, DS-Net has three appealing properties: 1) Strong backbone design. DS-Net adopts the cylinder convolution that is specifically designed for LiDAR point clouds. The extracted features are shared by the semantic branch and the instance branch which operates in a bottom-up clustering style. 2) Dynamic Shifting for complex point distributions. We observe that commonly-used clustering algorithms like BFS or DBSCAN are incapable of handling complex autonomous driving scenes with non-uniform point cloud distributions and varying instance sizes. Thus, we present an efficient learnable clustering module, dynamic shifting, which adapts kernel functions on the fly for different instances. 3) Extension to 4D prediction. Furthermore, we extend DS-Net to 4D panoptic LiDAR segmentation by the temporally unified instance clustering on aligned LiDAR frames. To comprehensively evaluate the performance of LiDAR-based panoptic segmentation, we construct and curate benchmarks from two large-scale autonomous driving LiDAR datasets, SemanticKITTI and nuScenes. Extensive experiments demonstrate that our proposed DS-Net achieves superior accuracies over current state-of-the-art methods in both tasks. Notably, in the single frame version of the task, we outperform the SOTA method by 1.8% in terms of the PQ metric. In the 4D version of the task, we surpass 2nd place by 5.4% in terms of the LSTQ metric.

Position-Guided Point Cloud Panoptic Segmentation Transformer

“…SemanticKITTI. We compare our method with RangeNet++ [33] + PointPillars [23], LPSAD [32], KPConv [41] + PointPillars [23], Panoster [16], Panoptic-PolarNet [48], DS-Net [18], EfficientLPS [38], GP-S3Net [37], SCAN [46] and Panoptic-PHNet [24]. Table 1 shows comparisons of LiDAR panoptic segmentation performance on the SemanticKITTI test split.…”

“…Compared with LiDAR-based 3D semantic segmentation, LiDAR-based panoptic segmentation further segments foreground point clouds into different instances. Most previous top-tier works [48,18,38,37,46,24] start from this difference and produce panoptic predictions following a three-stage paradigm, i.e., first predict semantic results, then separate instances based on semantic predictions, and finally fuse the two results. This paradigm makes the panoptic segmentation performance inevitably bounded by semantic predictions and requires cumbersome post-processing.…”

“…Backbone. In this work, we choose Cylinder3D [49] as our backbone for feature extraction due to its good generalization capability for panoptic segmentation [18].…”

4D-StOP: Panoptic Segmentation of 4D LiDAR using Spatio-temporal Object Proposal Generation and Aggregation

“…With the availability of RGB-D sensors, we have made progress on tasks like [16,17,27,40], 3D instance segmentation [11,14,15,22,26,54] and 3D object detection [35,41,59] which work on indoor scenes. Similarly, the accessibility to modern LiDAR sensors has made it possible to work with outdoor 3D scenes, where again recent works have targeted the tasks of 3D object detection [43,44,53,55], semantic segmentation [33,47,57,61], panoptic segmentation [5,21,34,60] and multi-object tracking [3,12,50,56]. Perceiving outdoor 3D environments from LiDAR data is particularly relevant for robotics and autonomous driving applications, and hence has gained significant traction in the recent past.…”

Input-Output Balanced Framework for Long-tailed LiDAR Semantic Segmentation

“…LiDAR semantic segmentation that could provide a thorough scene understanding, has attracted extensive studies. Most existing works usually pay much attention on point-to-structure representations, including spherical projection [1,2], bird-eye-view projection [3] and 3D voxelization [4,5], and another group of methods focus on the network architecture design [6,4,7]. However, these methods often neglect the inherit difficulty caused by the long-tailed data distribution.…”