LaserMix for Semi-Supervised LiDAR Semantic Segmentation

Abstract

Densely annotating LiDAR point clouds is costly, which restrains the scalability of fully-supervised learning methods. In this work, we study the underexplored semi-supervised learning (SSL) in LiDAR segmentation. Our core idea is to leverage the strong spatial cues of LiDAR point clouds to better exploit unlabeled data. We propose LaserMix to mix laser beams from different LiDAR scans, and then encourage the model to make consistent and confident predictions before and after mixing. Our framework has three appealing properties: 1) Generic: LaserMix is agnostic to LiDAR representations (e.g., range view and voxel), and hence our SSL framework can be universally applied. 2) Statistically grounded: We provide a detailed analysis to theoretically explain the applicability of the proposed framework. 3) Effective: Comprehensive experimental analysis on popular LiDAR segmentation datasets (nuScenes, SemanticKITTI, and ScribbleKITTI) demonstrates our effectiveness and superiority. Notably, we achieve competitive results over fully-supervised counterparts with 2× to 5× fewer labels and improve the supervised-only baseline significantly by 10.8% on average. We hope this concise yet high-performing framework could facilitate future research in semi-supervised LiDAR segmentation. Code will be publicly available 2 . * Equal contribution. 2 https://github.com/ldkong1205/LaserMix.Preprint. Under review.

Scribble-Supervised LiDAR Semantic Segmentation

“…dataset. We follow Kong et al [17] and generate a semisupervised dataset by uniformly sampling frames. As seen in Tab.…”

“…WS3D [20] utilizes region-level boundary awareness and instance discrimination to improve indoor and outdoor 3D semantic segmentation with simulated weak labels. Furthermore for semi-supervised learning, DiAL [32] uses a simple MT setup, GPC [16] proposes using a pseudo-label guided point contrastive loss, SSPC [8] utilizes self-training and LaserMix [17] uses a mixing operation to bring supervision to unlabeled frames. CPS [7] utilizes a Siamese structure to induce cross supervision.…”

“…Mix-Match [3] applies the same mixing operation to labeled and unlabeled frames in a semi-supervised setting while generating labels via guessing and sharpening for unlabeled parts to provide supervision. Specifically for semi-supervised learning on LiDAR point clouds, LaserMix [17] aims to introduce variability through cylindrical and range-view partitioning and mixing.…”

A survey on weakly supervised 3D point cloud semantic segmentation

“…Zhang et al [27] proposed a method for knowledge transfer using self-supervised point cloud colourisation based on RandLA-Net. Kong et al [36] argued that LiDAR scans contain rich spatial priors and proposed LaserMix, a method for combining scans from separate LiDAR laser beams.…”

“…Using self-supervised learning and label propagation, Zhang et al [27] restricted and optimised the network; Cheng et al [28] added pseudo label and attentional feature constraints; Wang and Yao [30,33], Liu et al [31,43], Shi et al [35] and Lu et al [41] all incorporated consistency and pseudo label constraints. As a novel hybrid constraint, Kong et al [36] merged pseudo labelling with weakly supervised information. Liu et al [40] proposed multiple label constraints and combined them with multiple examples learning for supervised training.…”