LiDARNet: A Boundary-Aware Domain Adaptation Model for Point Cloud Semantic Segmentation (2003.01174v3)

Abstract: We present a boundary-aware domain adaptation model for LiDAR scan full-scene semantic segmentation (LiDARNet). Our model can extract both the domain private features and the domain shared features with a two-branch structure. We embedded Gated-SCNN into the segmentor component of LiDARNet to learn boundary information while learning to predict full-scene semantic segmentation labels. Moreover, we further reduce the domain gap by inducing the model to learn a mapping between two domains using the domain shared and private features. Additionally, we introduce a new dataset (SemanticUSL\footnote{The access address of SemanticUSL:\url{https://unmannedlab.github.io/research/SemanticUSL}}) for domain adaptation for LiDAR point cloud semantic segmentation. The dataset has the same data format and ontology as SemanticKITTI. We conducted experiments on real-world datasets SemanticKITTI, SemanticPOSS, and SemanticUSL, which have differences in channel distributions, reflectivity distributions, diversity of scenes, and sensors setup. Using our approach, we can get a single projection-based LiDAR full-scene semantic segmentation model working on both domains. Our model can keep almost the same performance on the source domain after adaptation and get an 8\%-22\% mIoU performance increase in the target domain.

• The paper introduces a dual-branch model that leverages Gated-SCNN to incorporate boundary information for accurate segmentation.
• It preserves source domain performance while achieving an 8%-22% mIoU increase on target domains across diverse LiDAR datasets.
• The integration of the SemanticUSL dataset and robust adaptation methods offers significant benefits for autonomous vehicles, robotics, and mapping applications.

LiDARNet: A Boundary-Aware Domain Adaptation Model for Point Cloud Semantic Segmentation

The paper under discussion, "LiDARNet: A Boundary-Aware Domain Adaptation Model for Point Cloud Semantic Segmentation", introduces a model designed to address the challenges associated with LiDAR scan full-scene semantic segmentation. This model is particularly notable for its boundary-aware domain adaptation capabilities, facilitating improved interpretation and mapping between different data domains in various LiDAR point cloud applications.

Framework and Methodology

The authors present a model named LiDARNet, which employs a dual-branch structure to concurrently extract domain-specific and shared features. This architecture contributes to the segmentation process by fostering the extraction of features that are critical for distinguishing domain characteristics, while also identifying commonalities across differing data domains.

Central to LiDARNet is the integration of Gated Shape Convolutional Neural Networks (Gated-SCNN) into its segmentor component. This choice underscores the model's commitment to leveraging boundary information, effectively addressing the edge delineation challenge that is often encountered in semantic segmentation tasks involving LiDAR data. By utilizing these advanced techniques, the model enhances its prediction accuracy for full-scene semantic segmentation labels.

Dataset and Experimental Design

The research introduces SemanticUSL, a novel dataset developed for domain adaptation in LiDAR point cloud semantic segmentation. It is aligned with the data format and ontology of the existing SemanticKITTI dataset, allowing for seamless benchmarking and evaluation. SemanticUSL, together with real-world datasets like SemanticKITTI and SemanticPOSS, provides a diverse testing environment due to variation in channel distributions, reflectivity distributions, scene diversity, and sensor setups.

Results and Implications

Empirical evaluations demonstrate that the model preserves near-original performance on the source domain after adaptation and achieves a significant performance boost — an 8%-22% increase in mean Intersection over Union (mIoU) — for the target domain. These results indicate the efficacy of the proposed method in maintaining accuracy while also enhancing domain transfer capabilities.

The implications of this research are substantial for practical applications in robotics, autonomous vehicles, and geographic information systems, where LiDAR data is extensively used. By minimizing the domain gap, this methodology enhances model robustness and adaptability across different operational environments.

Conclusion and Prospects

LiDARNet exemplifies a progressive step in boundary-aware domain adaptation for LiDAR scan segmentation, demonstrating both theoretical and practical advancements. The introduction of the SemanticUSL dataset further enriches the landscape for future studies in semantic segmentation. Looking forward, extending this framework to incorporate additional modalities and exploring unsupervised learning techniques stand as promising avenues for subsequent research, potentially driving further improvements in semantic segmentation accuracy and generalizability across various domains.