Collaborative Learning for Semi-Supervised LiDAR Semantic Segmentation

Annotating large-scale LiDAR point clouds for 3D semantic segmentation is costly and time-consuming, which motivates the use of semi-supervised learning (SemiSL). Standard LiDAR SemiSL methods typically adopt a two-step training paradigm, where pseudo-labels are separately generated from a single distillation source, either from the same or another LiDAR representation. Such supervision relies on a unique source of pseudo-labels, which can reinforce confirmation bias and propagate errors during training, ultimately limiting performance. To address this challenge, we introduce CoLLiS, a novel framework that leverages Collaborative Learning for LiDAR Semi-supervised segmentation. Unlike prior paradigms with decoupled pseudo-labeling and training phases, CoLLiS trains multiple representations collaboratively in a single step by treating them as coequal students. Each student is adaptively distilled from multiple representations, while inter-student disparities are monitored online to resolve contradictory supervision and effectively mitigate confirmation bias. Extensive experiments on three datasets demonstrate that CoLLiS consistently outperforms state-of-the-art LiDAR SemiSL methods, with particularly strong gains in low-label regimes.

Self-driving cars and robots use LiDAR sensors to understand the 3D world around them, but training AI systems to label every object in LiDAR scans usually requires a large amount of expensive human annotation. This paper studies how to train LiDAR segmentation models when only a small fraction of the data is labeled and the rest is unlabeled. Existing methods often rely on one model or one view of the LiDAR data to create training labels for the unlabeled scans, which can cause the system to reinforce its own mistakes. We propose CoLLiS, a collaborative learning framework in which several models look at the same LiDAR scene through different representations and teach each other during training. Instead of trusting one source blindly, CoLLiS compares the agreement and confidence of the models to decide which pseudo-labels are reliable. It also adjusts data augmentation during training, making the learning task harder or easier depending on how much the models agree. Experiments on three public LiDAR benchmarks show that CoLLiS improves performance over previous semi-supervised methods, especially when labeled data is scarce. This can reduce the annotation burden for building reliable 3D perception systems in autonomous driving and robotics.