People - Research Staff / Wei Hao

IRIS Research Student

Wei Hao

Ph.D. Student

		Office:	209 Ferris Hall The University of Tennessee Knoxville, TN 37996
		Telephone:	(865) 974-9737
		Fax:	(865) 974-5459
		E-mail:
		Personal Web Page:

Current Research Work

3D Reconstruction Based on Two Wide Baseline Stereo Frames

Recovery of the lost dimension information during the course of imaging is one of the central problems in computer vision. Stereo is an attractive technique for depth perception from multiple views. Small baseline stereo have been investigated in lab and real world applications by many researchers, while there exist certain problems for wide baseline applications.

The two main problems we target in this research is:

1. Wide baseline sparse feature point matching.

2. stereo image Rectification.

The diagram we proposed is shown in Figure 1.

1. Wide baseline sparse feature point matching

The crucial processing that is needed to weakly calibrate the stereo system is finding out some accurate matching between reliable sparse point features. We proposed a novel algorithm which combines the global geometric information with the powerful local features detected by SIFT/MSER/Harris-Affine etc. The experimental results show enhancement over the current state of arts.

Wide baseline Image pair of a library. (courtesy of Geometric Vision Group of Oxford University)

Matching Results generated by applying SIFT descriptor matching only.

In the left image. 41 matches with 6 bad matches (error match rate: 14.6%).Notice that the wrong matches caused by repeated patterns (marked by red circle) and symmetry (marked by yellow circle).

In the right image. 101 Matching with 3 bad matching (error match rate 2.9%).Bad Matching also caused by repeated patterns but the error matches are near, which should be caused by the reinforce of the structure of the point sets.

Wide baseline Cottage Image pair courtesy of Oxford Geometric Vision Group.

Results:

In the left image. SIFT can Only generate 2 good matches out of ten.

In the right image. Our proposed method generates 54 matches and most of them are good.

2. Image Rectification

Image Rectification can reduce the 2D dense match to a 1D scan line search, thus make the computation cost down greatly.

Based on a robust sparse matching. We also use Robust epipolar geometry estimation method. We proposed a novel algorithm which can be applied in weakly calibrated cases.

The results of our algorithms on the cottage image pair is shown in the following Figure.

Schedule:
Monday-Friday: 8:30am-5:30pm

Please contact Webmaster if you have any questions.