We filmed moving snowflakes in front of a tainted wall at a high frequency (210Hz), and the points were detected using very simple means (thresholding and a automatic clean-up, see our accompanying paper) and the ground-truth was then recovered by hand on this simple to track version. The snow sequence has then been subsampled to 30Hz by keeping one frame out of every 7, to obtain a challenging to track point clouds sequence.
Feel free to use this sequence in your publications, but please cite the accompanying paper as a reference to its origin.
Here are the results for each algorithm (astre-noholes and astre-holes) for the standard parameter ε = 0 and the optimal parameter ε = 5 in terms of recall and precision (computed on the number of correct links found) and the number of found trajectories as a reference:
algorithm | recall | precision | # trajs |
astre-noholes, ε = 0 | 0.50 | 0.99 | 39 |
astre-holes, ε = 0 | 0.55 | 0.91 | 34 |
astre-noholes, ε = 5 | 0.71 | 0.95 | 124 |
astre-holes, ε = 5 | 0.76 | 0.92 | 91 |