Reproducing SMURF baseline numbers on TAP-Vid DAVIS

[SeKim12]

Hi, thank you for introducing this method and the implementation! I had a quick question regarding the table in your paper with TAP-Vid metrics for different baseline models (page 5, Table 1). Using the official SMURF implementation and provided checkpoint, I am unable to reproduce the numbers that were reported here. Here is a side-by-side comparison

Metrics	Paper	My code
AJ	29.8	25.7
AD	27.4	32.4
Avg pts	42.7	37.3

I am resizing all DAVIS frames (and labels) to (256, 256) when computing the metric. I am also using the sintel-smurf checkpoint. It would be immensely helpful if you could point me to the specific configurations or code that you used to generate the numbers reported in the paper. I found the results to be quite interesting, especially because I could not find any other papers evaluating smurf on Tap-Vid DAVIS, and would love to look into this further. Thank you!