ā† Back to Projects

Learning-Based Point-Cloud Refinement via Hole Filling

Model to detect and fill holes generated due to unnatural reasons

Method
Synthetic hole generation creates training examples for a learned model that detects and fills gaps in 3D point clouds.
Stack
Python, PyTorch, PyTorch3D, Open3D

Developed a deep learning model to identify and repair point cloud gaps caused by occlusion and reflections. Designed an algorithm to generate datasets with realistic holes in point clouds for training and testing purposes.

Architecture

Block diagram of the point-cloud hole-filling model
Block diagram of the project.

Dataset generation

Dataset-generation block diagram
Block diagram for generation of the dataset.

Results

Point-cloud hole-filling results
Results.