Profile

Chinese Name:黃柏豪
English Name:Po-Hao Even Huang
Gender:Male
Job:MediaTek Inc.
Grades: PhD Students
Graduated:2010

Contact

Email:Even@cs.nthu.edu.tw

Reserch

Group in CVLAB: 3D group
Interested in:3D Reconstruction, Image Processing
Thesis:Silhouette-Based Camera Calibration and 3D Shape Recovery from Mirror Reflection or Circular Motion
3D modeling from images is a popular research topic in computer vision and computer graphics. During the 3D reconstruction process, the camera calibration step plays an important role that estimates camera parameters from the information extracted from images. Traditional feature-based reconstruction algorithm, i. e. the Structure from Motion (SfM) algorithm, exploits the feature correspondences across views for camera calibration. However, to robustly estimate correspondences is a challenging problem, especially for wide-baseline images or texture-less objects. Instead of making use of feature correspondences, in this thesis, we propose to exploit the object silhouettes for camera calibration without identifying any feature points in advance or using any specific calibration pattern.
From an arbitrary moving camera, the captured image silhouettes of an object can not provide enough constraints for estimating camera parameters without a good initial guess; therefore, we need to restrict the camera motion that makes camera calibration from silhouettes possible. We hereby propose two 3D modeling algorithms based on different scene settings; namely the Shape from Mirror Reflection (SfMR) and the Shape from Circular Motion (SfCM) algorithms. In the SfMR algorithm, an object is put in front of planar mirrors, and the camera positions are constrained by mirror reflection property. In the SfCM algorithm, an object is put on a turn-table, and the camera positions are constrained by the circular motion property. From different properties, we derive constraints to calibrate the camera and reconstruct the 3D shape of a model using only the object silhouettes. Experimental results on both synthetic and real data sets demonstrate the robustness of the proposed algorithms and their performance.
even