Image Processing Lab (IPL)
Point cloud registration with low overlapping
keywords DEEP LEARNING, VIDEO ANALYSIS
Reference persons ENRICO MAGLI
Thesis type RESEARCH
Description An important problem in 3D computer vision is point cloud registration. In general, a point cloud is a set of 3d points that can represent an object or a scene. Given a pair of point clouds the goal of point cloud registration is to find the rigid transformation that aligns the two data. This is a fundamental problem for many challenging applications, such as 3D reconstruction, motion estimation or object detection.
Several methods that tackle this problem are present in the literature, but they are mainly focused on point clouds entirely overlapped or with a high percentage of overlapping points. Therefore, the low-overlap scenario, that is important for real applications, is still unexplored. The purpose of this thesis is to investigate an innovative deep learning neural network that address point cloud registration using state-of-the-arts techniques such as transformers and rotation-invariant and equivariant features extraction. The student will develop the network and make various experiments, starting from some preliminary results.
Required skills Candidate students should have some background on neural networks. Some experience of TensorFlow environment and Python programming are desirable, along with good programming skills.
Deadline 12/05/2023 PROPONI LA TUA CANDIDATURA