A self-organization based optical flow estimator with GPU implementation

This work describes a parallelizable optical flow field estimator based upon a modified batch version of the Self-Organizing Map (SOM). This estimator handles the ill-posedness in gradient-based motion estimation via a novel combination of regression and self-organization. The aperture problem is treated using an algebraic framework that partitions motion estimates obtained from regression into two sets: one (set H c ) containing motion estimates with high confidence and another (set H p ) with low confidence. The self-organization step uses a uniquely designed pair of sets: training set ( Q = H c ) and the initial weights set ( ). It is shown that with this specific choice of training and initial weights sets, the interpolation of flow vectors is achieved primarily due to the regularization property of the SOM. Moreover, the computationally involved step of finding the winner unit in SOM simplifies to indexing into a 2D array making the algorithm highly scalable. To preserve flow discontinuities at occlusion boundaries, we have designed an anisotropic neighborhood function for SOM that uses a novel optical flow constraint equation residual-based distance measure. A multi-resolution or pyramidal approach is used to estimate large motion. As self-organization based motion estimation is computationally intense, parallel processing on graphics processing units is used for speedup. As the algorithm is data (rather datum) parallel, with sufficient number of computing cores, the implementation of the estimator can be made real time. With the available ground truth from Middlebury database, error metrics like average angular error and average end point error are computed and are shown to be comparable with other leading techniques.