Multi-view Dense Point Correspondence Ground-Truth Dataset

This is a dataset consisting of 13 calibrated views of the same scene as well as dense point correspondence maps between each pairs of views. The dense correspondence is obtained using a probabilistic structured light framework, which yields one or more correspondence options for each pixel in a given view, along with confidence scores associated with each of these options. The confidence scores are based on observations of the scene under both ambient and structured light. Calibration is obtained by tracking 40 points across all 13 views and using bundle adjustment. Calibration files consist of K (intrinsic camera matrix), R (rotation matrix) and T (translation vector) respectively. Maximum reprojection error is measured to be 1-2 pixels in the full-resolution sequence and subpixel for half and quarter sized versions. Also made available here is a 3D point cloud of the scene, reconstructed using one of the correspondence maps and calibration information.

 

The technical report for this project is currently under review and will be made available after August 15th, 2012.

 

We gratefully acknowledge the support of NSF Grant 1116140.

 

 

 

 

 

 

 

 

FULL SIZED (4272x2848)

Images

Calibration

Dense Correspondence

40 Manually Marked Points

3D Point Cloud

 

HALF SIZED (2136x1424)

Images

Calibration

Dense Correspondence

 

QUARTER SIZED (1068x712)

Images

Calibration

Dense Correspondence