Deep Closest Point Reproduction
Embedded Files
Left: an (un)aligned desk. Center: an (un)aligned light. Right: an (un)aligned airplane.
Clockwise from top left: input, results of DCP (ours), results of ICP initialized with our DCP, results of ICP without good initialization
Abstract
This project reproduces Deep Closest Point, a promising learning-based method for the point cloud registration problem. Our implementation shows less translation error than, and comparable rotation error with trained-from-scratch official implementation. This project was one of the top-scored final project for KAIST 2021 spring CS492(H) Special Topics in Computer Science<Machine Learning for 3D Data> course. (2 out of 27 students)
Contribution: I was solely responsible for this project.
Framework
Page updated
Google Sites
Report abuse