기계공학부 구성원들의 많은 관심과 참여 부탁 드립니다.

▣ 주   제: Success of Structure-from-Motion and Beyond

▣ 연   사: 홍제형 교수

▣ 소   속: 한양대학교 융합전자공학부

▣ 일   시: 2022. 9. 1.(Thu) 17:00

▣ 장   소: 제1공학관 N404호 

▣ 초   청: 이준상 교수

▣ 초   록

Structure-from-motion (SfM) is one of the most successful pipelines ever built in computer vision, and is widely used as a benchmark and/or preprocessing tool for neural 3D reconstruction tasks. The optimization problem it solves is nonlinear, which many still believes is the major difficulty in SfM and the reason to go for an incremental multi-stage pipeline that breaks down a difficult optimization problem into easier subtasks. For the first part of the talk, I will present some earlier works demonstrating that the actual optimization problem solved in SfM may not be as difficult-to-solve as previously perceived if input data is clean, implying the success of SfM lies in its ability to consistently discard anomalous measurements and manipulate data matrix. Largely inspired by the findings of this work, the second part of the talk comprises summary of recent work in which an incremental pipeline is devised for a completely different problem of 3D pot reassembly. The talk will conclude by presenting my current research efforts in SfM addressing end-users’ privacy concerns.