Pixels to Points™ Information

The Pixels-to-Points tool identifies matching location across a series of overlapping images, and then use those matching features to reconstruct the 3-dimensional structure of the landscape and features within it using Structure from Motion (SfM).

First the tool calculates the focal length of each image if it was not in the camera metadata. Next it identifies feature points of distinguishing features that can be located across images, and then locates those feature points across all images and correlates the relationship between matching pairs.

With this information, the Structure from Motion algorithm calculates a sparse point cloud using 3D point triangulation. This calculates a set of 3D points from the feature points, using either the incremental method or the global method.

The sparse point cloud is geographically registered using the GPS data from the image metadata, and any specified ground control points. Finally, Multi-View Stereovision is performed to densify the sparse point cloud and produce the final output point cloud.

The calculations used in the Pixel-to-Point tool involve a few stages of randomization in the process of the calculation. This is done in order to narrow down appropriate parameters for the algorithms, as well as to reduce the amount of computation, however as a result the tool will not generate exactly the same point cloud each time it is run.

Analysis Method

There are two types of Analysis Methods that can be used to build the key feature points, the Incremental method and the Global method.

The Incremental method starts from two images, and progressively adds more, recalculating the parameters and locations of the points to minimize the error. The Global method considers the keypoints across all images at the same time.

The Incremental method is recommended for images with lower overlap. However it can take longer to process large sets of images, since it is a progressive method. It can also have drift effects, where the residual errors are unevenly distributed across the point cloud because of the incremental consideration of feature points.

The Global method is recommended for large datasets with high amounts of image overlap between images and many angles of image orientation.