Detial config for different dataset

[haowang020110]

Hi, author
Thanks for your kindly share the code first!
I try to reproduce the results in the paper, using the C3VD dataset sequence cecum_t1_a, which seems to be the Fig2 (b) results.

I use the default slam config by simply run python run_slam.py --data_root $data_root --output_folder $experiment_folder --name cecum_t1_a .
The image is undistorted.
The results is worser compared to the paper.

I guess it may raise from different experimental set up?

[TMT22]

From the looks of it, I am guessing some of the point tracks are inconsistent with the camera motion in your case.

If you are not supplying a mask, it would be good to do so to filter out some non-moving tracks at the edges. Been a while, but if I remember correctly the distortion for this dataset was quite strong (as in, even after undistortion the border of the image was not that good), so ideally features closer to the center, where the undistortion is good, should be tracked. Also, I believe that we used "R2D2" to sample the points for this trajectory, but the default in run_slam is "Uniform".

Note that by this point there are also follow-ups to the CoTracker version used here out there, these will likely lead to better performance as well.