Inevitable Identity Switches - Help with post-processing?

[djg9324]

I've been working on building a reliable multi-animal model using bottom-up processing. The tracking works fine and the animals (two mice) are being identified, but I seem to have identity switches throughout moments in the video files (especially when mice get close to one another) no matter how many frames I adjust and train. I am wondering if there is a post-processing code that assists in enforcing identity consistency. I have written something but am unsure if it's working as the cell runs for a long time with no updated progress.

*I understand the ID models are made to help with identity switches, but our GUI lacks the memory to run ID

[gitttt-1234]

Hi @djg9324 !!

Could you please share the logs when you run inference/ tracking? The logs will have the set of parameters used to initialize the inference pipeline. Could you also try using the normalized_instance similarity method? You could do this by adding this --tracking.similarity normalized_instance to the sleap-track cli.

Thanks,

Divya

[djg9324]

Ah okay, I haven't tried the normalized_instance method. I've been using Instance (is normalized instance new? I haven't noticed that one before). I'll try that real quick. Meanwhile, these are my parameters for training


And my parameters for Inferring

[djg9324]

I just tried with normalized instance, but identity switches are still occurring.