Semester of Graduation
Fall 2023
Degree
Master of Science in Computer Science (MSCS)
Department
Computer Science
Document Type
Thesis
Abstract
3D point cloud segmentation segments the 3D point cloud data into different regions/instances depending on their features that have numerous applications in robotics, autonomous driving, digital twinning, augmented reality, etc. The majority of the existing point cloud segmentation methods depend on class labels to identify 3D objects in the surroundings. Our work focuses on segmenting point clouds into different regions/instances in an object-agnostic manner for any number of objects in the environment. Given the point cloud, our method can segment the entire scene into multiple instances without depending on object shape and size. We leverage the power of the self-attention mechanism combined with the region-growing approach to completely segment an environment/area ranging from small-scale to large-scale scenes. We used vector attention instead of scalar dot product attention for more accurate results. The usage of self-attention in local neighborhoods helps to capture the local context of neighborhood points and encodes the positional information in the network. Our novel approach outperforms previous class-agnostic and class-specific segmentation methods by substantial margins.
Date
10-18-2023
Recommended Citation
Gyawali, Dipesh, "LRTransformer: Learn-Region Transformer for Object-Agnostic Point Cloud Segmentation" (2023). LSU Master's Theses. 5856.
https://repository.lsu.edu/gradschool_theses/5856
Committee Chair
Zhang, Jian