Unsupervised Domain Adaptation for 3D LiDAR Semantic Segmentation

Semantic understanding of large-scale 3D environments from LiDAR point clouds is a key requirement for autonomous systems such as robotics, intelligent vehicles, and mapping applications. Although recent deep learning approaches achieve high accuracy on benchmark datasets, their performance degrades significantly when applied to new environments due to domain shift. These shifts arise from differences in sensor configurations, point density, acquisition platforms, and geographical variations. At the same time, annotating 3D point clouds remains expensive and time-consuming, limiting the scalability of supervised approaches.

In this work, we address this challenge by proposing a novel unsupervised domain adaptation (UDA) framework for 3D LiDAR semantic segmentation. The goal is to transfer knowledge from a labeled source domain to an unlabeled target domain without requiring additional annotations. Our approach follows a two-stage training strategy that combines contrastive representation learning with a robust multi-model pseudo-labeling scheme.

In the first stage, we perform unsupervised contrastive pre-training to learn domain-invariant feature representations. The input point clouds are first decomposed into structural segments by removing the ground plane using RANSAC and clustering the remaining points via DBSCAN. Two augmented views of the same scene are generated through geometric and point-level transformations. These views are processed by a shared backbone network, and segment-level features are extracted and optimized using a contrastive loss. This encourages consistent representations for corresponding segments while separating unrelated structures, enabling the model to capture robust geometric and semantic patterns without manual supervision.

In the second stage, we introduce a multi-model pseudo-labeling strategy to generate reliable supervision for the target domain. Instead of relying on a single segmentation model, we leverage an ensemble of diverse state-of-the-art architectures, including projection-based, voxel-based, and hybrid approaches. Each model produces predictions on the target data, which are then aggregated using a hard voting mechanism. This ensemble-based strategy reduces the noise and bias associated with individual models and results in high-quality pseudo-labels.

The pre-trained network is subsequently fine-tuned on the target domain using these refined pseudolabels as supervisory signals. This allows the model to adapt effectively to new sensor characteristics and environmental conditions, improving its generalization performance.

We evaluate the proposed framework on challenging cross-domain scenarios, adapting from the SemanticKITTI
dataset to unlabeled target datasets such as SemanticSlamantic and SemanticPOSS. The results demonstrate significant improvements over direct transfer and single-model adaptation approaches. In particular, the ensemble pseudo-labeling strategy leads to notable gains in segmentation accuracy and robustness, especially for difficult and small object classes such as pedestrians and traffic signs. Qualitative results further highlight the effectiveness of the method, showing cleaner and more consistent segmentation outputs compared to individual models.

Overall, this work demonstrates that combining contrastive feature learning with ensemble-based pseudolabeling provides an effective solution for unsupervised domain adaptation in 3D LiDAR semantic segmentation. The proposed framework successfully bridges domain gaps caused by sensor and environmental variations without requiring target domain annotations, making it suitable for real-world deployment in autonomous systems.