Scribble-Supervised LiDAR Semantic Segmentation (2203.08537v2)

Abstract: Densely annotating LiDAR point clouds remains too expensive and time-consuming to keep up with the ever growing volume of data. While current literature focuses on fully-supervised performance, developing efficient methods that take advantage of realistic weak supervision have yet to be explored. In this paper, we propose using scribbles to annotate LiDAR point clouds and release ScribbleKITTI, the first scribble-annotated dataset for LiDAR semantic segmentation. Furthermore, we present a pipeline to reduce the performance gap that arises when using such weak annotations. Our pipeline comprises of three stand-alone contributions that can be combined with any LiDAR semantic segmentation model to achieve up to 95.7% of the fully-supervised performance while using only 8% labeled points. Our scribble annotations and code are available at github.com/ouenal/scribblekitti.

• The paper presents a scribble-based annotation method that labels only 8% of points, achieving up to 95.7% of fully-supervised performance.
• It leverages a Mean Teacher model with partial consistency loss and class-range-balanced self-training to address LiDAR data’s long-tailed distribution.
• The introduction of Pyramid Local Semantic-context enriches feature extraction, enhancing pseudo-label quality during semantic segmentation.

Scribble-Supervised LiDAR Semantic Segmentation

The paper "Scribble-Supervised LiDAR Semantic Segmentation" introduces a novel approach for efficiently annotating LiDAR point clouds using scribbles, aimed at reducing the annotation cost associated with dense labeling. The researchers propose ScribbleKITTI, a new dataset featuring scribble annotations for LiDAR semantic segmentation, which allows significant annotation time savings while maintaining high model performance compared to fully-supervised methods.

Overview and Methodology

The paper highlights the challenges posed by the growing volume of LiDAR data, especially for applications like autonomous vehicles, where dense annotation for semantic segmentation can be prohibitively time-consuming and expensive. The proposed approach leverages scribble annotations, which are less detailed than full annotations, to label only 8% of the points in the dataset. The authors introduce a pipeline to minimize performance degradation typically associated with weak supervision.

The methodology includes three primary contributions:

1. Mean Teacher with Partial Consistency Loss: The authors adapt the mean teacher model for LiDAR data by applying a consistency loss only to the unlabeled points, thereby enhancing model training without distorting the supervision of the labeled points by the teacher network's uncertainties.
2. Class-Range-Balanced Self-Training (CRB-ST): To tackle the bias introduced by the inherent long-tailed distribution of LiDAR scenes, the authors propose a class-range-balanced self-training scheme. This involves generating pseudo-labels by sampling predictions with confidence thresholds while ensuring balanced representation across both classes and spatial ranges within the LiDAR data.
3. Pyramid Local Semantic-context (PLS): The introduction of a novel descriptor enriches the data by leveraging a semantic prior, computed as class distributions over spatial bins at multiple resolutions. This augmentation enhances feature richness and aids in producing higher-quality pseudo-labels during self-training.

Experimental Results

The experiments conducted on the SemanticKITTI dataset demonstrate the efficacy of the proposed methods. The results indicate that the scribble-based approach achieves up to 95.7% of the performance of fully-supervised training, showcasing the potential of scribble annotations in significantly reducing the annotation overhead without substantial performance compromise. The researchers provide compelling comparisons across different state-of-the-art baseline models, including Cylinder3D, MinkowskiNet, and SPVCNN, reinforcing the general applicability of the proposed workflow.

Implications and Future Directions

The implications of this research are broad, both practically and theoretically. Practically, the adoption of scribble annotations and efficient supervision strategies can lead to substantial cost savings and faster data preparation cycles for 3D LiDAR datasets. Theoretically, this work opens new avenues for exploring efficient weak supervision techniques in structured 3D environments, potentially enhancing the scalability of LiDAR-based sensing in dynamic real-world applications.

The paper hints at possible future work involving applying this methodology to other datasets and exploring more sophisticated weak supervision techniques that could further bridge the performance gap with fully-supervised models. The public availability of the ScribbleKITTI dataset and the associated code fosters further research and innovation in this direction.

In summary, this paper presents a thoughtful approach to addressing the bottleneck of LiDAR data annotation, paving the way for more efficient methods in 3D semantic segmentation. The combination of scribble annotations with sophisticated self-training frameworks offers a valuable contribution to the field of computer vision and autonomous systems.