Presenter(s)

Nordin Abouzahra

Comments

3:00-4:15, Kennedy Union Ballroom

Download

Download Project (3.2 MB)

Description

Modern 3D sensing technologies like LiDAR generate expansive point clouds, necessitating efficient data annotation techniques. Traditional methods focus on a per-point approach, discriminating individual points via global features such as depth or other available sensor data. While functional, this approach becomes unwieldy as point clouds grow in size. To address this challenge, we introduce the use of superpoints as part of the annotation process. By leveraging superpoints, we can exploit geometric information within the point cloud that would otherwise be overlooked, thereby assisting the annotator. These superpoints represent clusters of locally coherent data, offering a more interpretable and manageable unit for labeling compared to individual points. Our approach begins by processing the point cloud into superpoints. This process produces superpoints that belong to a partition level, ranging from a fine-to-coarse representation. This structure lays the groundwork for more precise, human-guided refinement. Annotators can leverage the inherent spatial and geometric coherence highlighted by the superpoints to expedite their annotation alongside traditional methods. Additionally, annotators can choose the granularity at which to operate. For instance, they may start with a rough pass using the coarsest representation and then refine their labels at the finest level. We conducted comprehensive experiments across multiple large-scale point clouds to evaluate the benefits of our method. The results demonstrate a reduction in annotation time, accompanied by enhanced label precision. This improvement is attributed to the method’s capacity to convey complex spatial information through easily identifiable clusters, thereby reducing the cognitive load on human annotators.In conclusion, by redefining the annotation process with a focus on superpoints, our framework offers a robust solution for the challenges of large-scale 3D data labeling. This advancement not only improves procedural efficiency but also lays a foundation for more scalable and detailed annotation workflows in diverse 3D applications.

Publication Date

4-23-2025

Project Designation

Graduate Research

Primary Advisor

Vijayan K. Asari, Theus H. Aspiras

Primary Advisor's Department

Electrical and Computer Engineering

Keywords

Stander Symposium, School of Engineering

Institutional Learning Goals

Practical Wisdom; Community; Diversity

The Aesthetic Advantage: Enhancing Visual Clarity in Point Cloud Labeling with Superpoints

Share

COinS