About nfl helmet assignment kaggle competition

Got bronze model, that was fun to compete

Here I discuss short overview of my solution:

At first for each 20th frame I bruteforce homography matrix minimizing the euclidean distance to all players (I compute homography matrix by 4 extremal points in x/y). Then I receive tracks from deepsort and try to make assignment. For each frame I solve weighted assignment problem (by hungarian method) between tracks and trackings based on last k-frames. The cost is euclidean distance between these subtrajectories (actually I tried to use additional information - like orientation/speed/acceleration and use metric learning for these - but got minor improvements).

The results become much worse after player collision - because of id switches and tracks termination.