Building Computer Vision Projects with OpenCV 4 and C++

Finding feature tracks

The concept of feature tracks was introduced in SfM literature as early as 1992 in Tomasi and Kanade's work (Shape and Motion from Image Streams, 1992) and made famous in Snavely and Szeliski's seminal photo tourism work from 2007 for large-scale unconstrained reconstructions. Tracks are simply the 2D positions of a single scene feature, an interesting point, over a number of views. Tracks are important since they maintain consistency across frames than can be composed into a global optimization problem, as Snavely suggested. Tracks are specifically important to us since OpenCV's sfm module allows to reconstruct a scene by providing just the 2D tracks across all the views:

Having already found a pair-wise match between ...