a fun next step would be to run the SAM-Track demo on colab and upload a shot to it, skipping any fading frames at the start of the shot. have it mark every object in the first good frame a different color, then have it track them through the shot. [i wonder if results could be improved by picking different starting frames and running backward. for example, the shot could be decimated and a sequence of starting frames tried.] [there's a graph-like problem here where when objects are assembled or disassembled, they change in number. selecting the larger number here is the successful thing to do, but may take some engagement of the segmentation system.] [haven't looked into matching objects across shots. one idea is to just throw every object image into meshroom.]