Overlap of two mice leads to single detection and dataset creation error

[ttllzz]

I downloaded the software and successfully installed it using the program on GitHub. I also managed to run the demo with some AI modifications🙂. However, when using my own videos (two mice that occasionally overlap in body regions), the connected components are only recognized as a single object. This causes an error when I proceed to the second step, "Create training dataset." Moreover, the frame at which the error gets stuck varies each time (it seems the issue is that only one mouse is detected in certain frames, making it impossible to proceed).

My question is: while I could keep modifying the miss_object file with AI every time it fails, is there a better solution? Alternatively, should I adjust some parameters in the earlier steps to prevent this? Or are there specific settings I should tune during the initialization process to handle occlusions and overlaps more robustly?

[cabooster]

Hi,

Thank you for your feedback.

From your description, it seems that the issue may not be caused by the connected components detecting only one mouse in some later frames. In UDMT, during the tracking initialization step, the number of animals is determined manually in the first frame. For example, if two mice are selected at the beginning, the subsequent tracking process should keep the object number fixed by default, rather than changing the number of animals according to the connected components detected in each frame.

Therefore, even if two mice overlap in some frames, the program should not simply regard the video as containing only one animal because of temporary body-region overlap. The miss_object file is mainly used to handle difficult or missing detections, but normally it should not need to be repeatedly modified for this reason.

Could you please check whether both mice were correctly selected during the tracking initialization step, especially in the first frame?

Best,
Yixin Li