Using the validator

[K G]

Hey guys,

I improved the segmentation, such that idtracker.ai works better, even though it still enters protocol 3. Anways, I applied it to a video with 3 mice and length of 1 hour. The accuracy is at about 97 percent. 

Still I think that the identities change a lot and I dont know how to fix this with the validator. 

Please explain me how to do this. 

When 2 mice stand close to each other for a loger time, they create one big blob. This blob usualy has either one id or no id (see the attached screenshot). Is there a way to assign multiple Ids to this big blob?

Btw how do you calcualte the accuarcy? Is it for the accuarcy for tracking or for identitfying the mice?

[idtrackerai]

Hi,

If you mean you improved the segmentation code and you think your improvements could be beneficial for other users, I would be very happy to discuse and adding them in next release. If you mean you improved your recording and found the working segmentation parameters for your videos, then great!

Lots of details can cause the failure of Protocol 1 and 2. Some videos are very difficult to track for idtrackerai, even when they seem very easy for humans. Note that idtrackerai don't identify crossing blobs, it only identifies individual ones. Crossing blobs are solved assuming relatively short crossings and applying some simple logic. For videos where animals overlap for long periods of time, this crossing solver can perform poorly.

Once in the Validator, you can change identities by double clicking on any centroid and setting a new identity. Also, you can browse errors activating the interpolator for trajectory gaps (time slices when some identity is gone). The documentation can help (here).

The estimated accuracy is the average identification certainty of all the identified individual blobs. So, non identified blobs and crossing blobs don't matter regarding this metric.

Jordi