Short URL

Self Organizing Image Blocks

 

Other Photogrammetry areas:

Loading, please wait...

Self Organizing Image Blocks

Dr Jackson Cothren and students in GeoSciences and Computer Science and Engineering are using CAST facilties to experiment with a class of feature extractors and descriptors know as Scale and Affine Invarient Regional Descriptor (e.g. the SIFT operator) to determine if they can automonously orient large blocks of aerial imagery - without any a priori orientation knowledge - with enough precision to initialize traditional Automatic Aerial Triangulation (AAT) algorithms. Several approaches using Lowe’s SIFT operator, a multi-scale Harris corner detector, and the SURF operator, all described using Lowe’s method, have been implemented in MATLAB/C++ and tested against a variety of image types. Our experiments to date have demonstrated the effectiveness of this approach in image blocks with as many as 60 images (~5000 x ~7000 pixels per image, 6” GSD) in a 60% overlap / 30% sidelap collection. Matches are proposed by clustering features 128-dimension feature descriptor space and are pruned using RANSAC-based relative orientation. Enough multi-image (3+ ray) matches are found to completely estimate all exterior orientation parameters in a bundle adjustment with a standard error of less than 2 pixels at image scale. We are now experimenting with distributing this process across multi-node computers in the University of Arkansas' High Performation Computing Center.

  Jack Cothren Center for Advanced Spatial Technologies CAST University of Arkansas Integration of Terrestrial LiDAR and High-resolution Digital Images

Over 1100 point matches across two images. Overview of all matches are shown on the left with the two highlighted matches rendered as large magenta circles. Close-ups of the area surrounding the highlighted matches are shown in the middle (intensity images developed from the rgb originals). The far right pane shows the 128 dimensional feature vector of each point and the euclidean distance between the two vectors.