Issues with training a new classifier

[Divyansh Rana]

Hi everyone,

In JAABA, when I make a classifier for lunging,  when training, I labelled some positive instances as a lunge, after which I went ahead and predicted on the same set of videos used for training (JAABA automatically predicts as well after training) and I corrected predictions then by changing labels that classifier says as a positive lunge to none when its wrongly classifying. Now when I retrain and it repredicts it should ideally consider that "none" frame as none (since I labelled it that way) but instead it marks it as lunge only.

I dont understand why this is happening. Due to this if there are 30 frames of lunging my output of the classifier says 70 just because it marks frames within the none bracket as well after I trained it properly. Any ideas on this? I have attached a screenshot for reference.

[Divyansh Rana]

In reference to the trailing situation these "none errors" are what I am concerned about. I am getting none errors which are getting prominent as I add more training videos.

[Mayank Kabra]

Hi Divyansh,

I see that you have labeled a lot of data, which is great for learning an accurate classifier. However, the default number of iterations (100) might not be sufficient to train a sufficiently complex classifier. You should increase the number of training iterations to 200 or 500 and see how the new classifier performs (https://jaaba.sourceforge.net/Training.html#ClassifierParameters).

Another way to avoid mispredictions is to use postprocessing (https://jaaba.sourceforge.net/Training.html#PostProcessing). These would be effective at removing 1 or 2 frame incorrect lunge predictions.

HTH,

Mayank

--
You received this message because you are subscribed to the Google Groups "JAABA" group.
To unsubscribe from this group and stop receiving emails from it, send an email to jaaba+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/jaaba/ab281b99-06a1-42c0-8992-abb0a06ab63cn%40googlegroups.com.

[Pavan]

Hi Divyansh,

Mayank's suggestions on increasing training iterations and using post-processing are good starting points for addressing the "none" label bleed-through you are seeing.

On a related note, we have a validated JAABA lunge classifier developed as part of the DANCE pipeline (Yadav, Dey et al., eLife 2025). Direct transfer may be limited if your arena geometry and recording conditions differ substantially from ours, since JAABA features are sensitive to these parameters. That said, the GitHub repository (https://github.com/agrawallab/DANCE) and the publication describe our training procedure in detail, including how we handled class imbalance between "lunge" and "none" frames and how we selected training videos iteratively, which may provide useful guidance for building your own classifier.

Hope this helps.

Pavan