Refine tracklets vs labeled video


We are trying to track 1 mouse in a training box. Our videos are about 25 mins long (30 fps).

I split the video into 5 x 5mins videos and labeled about 50 frames for each videos for training.

When the training and evaluation is done, I analyzed a new 1 minute long video. The labeled video created from the evaluation step showed perfect tracking. However, when we go to the refine tracklets step, we see that the tracking is not optimal. For instance, tracking is perfect when the mouse is walking. Once the animal starts standing against the wall, the tracking start to diverge.

Screen cap from labeled video. The animal is walking along the wall.

Screen cap from refine tracklet. The animal is walking along the wall

Screen cap from labeled video. The animal is standing against the wall.

Screen cap from refine tracklets. The animal is standing against the wall.

Should we correct all the body parts that are out of place in the refine tracklets step? Any suggestions is great appreciated.

Thank you for creating this wonderful tool. It is a great way for us to look at animal behavior at a different dimension.



Hi, Allen:
Do you have to run this as a multi-animal project? If not, the refining is much easier in a single-animal project set up.

If you do have to run it as multi-animal, in the refine tracklets GUI, drag ALL of the markers in several frames. Don’t do all of the frames, but once you move a marker in one frame, make sure all of them get placed correctly.
Save, then back on the main GUI, before you do anything else, merge the data. That will take the frames you “edited”, and include only them in your labeled-data.

Alternatively, in the refine tracklets GUI, you can save without changing anything (saving is important, though). Then extract outliers for that video, and refine labels and merge (the workflow described for single animal projects in the docs).

Hi Brandon,

Thank you for the response. That makes me feel a lot better. I was not sure if it would also merge the uncorrected frames to the training data.

We do not have to run the multi-animal project. I simply followed the video tutorial and changed the individuals to individual1.

While waiting for someone to reply, I have refined quite a lot of frames. I will merge it and retrain. Thank you so much!



I have already