Object segmentation is a common task in many computer vision applications. The main purpose is that of identifying objects within an image separating them from their background. Semantic segmentation takes this approach one step further and tries to tag each object with a semantically relevant label (e.g. car, person, cyclist). Advances in deep learning have helped significantly. However, most object detection and semantic segmentation methods focus on 2D data (images). It is of interest to extend these approaches to other types of data, particularly 3D point clouds.
There are interesting challenges present in the point cloud domain that don’t exist in images. Such as point cloud registration and refinement of object boundaries. Some of the research activities required for this project include:
- Implementation of efficient algorithms to process the point clouds
- Development of point cloud registration algorithms
- Extension of semantic segmentation framework to the 3D domain