Advice for labelling datasets for maDLC 2.26


We’ve been working with maDLC 2.26 and are enjoying the performance upgrades – thank you for your hard work!

We are just expanding our datasets now and wanted a bit of advice.

We have a max of three subjects, or individuals, in a given frame but this varies across a given video as they move in and out of the picture. It’s relatively infrequent that they’re all in frame together.

Is it okay to label frames with just one or two subjects? or would the eventual model perform best if the dataset was created pretty much exclusively with frames where all individuals are in view?

Could you comment on how the number of individuals annotated in a given frame affects the performance of the eventual trained model?

Many thanks


Hi Sotiris, great to hear. whatever is the max # of individuals in the training images you label should be the # of individuals you list originally. But then that network can be used to analyze N individuals, simply just change the number in the inference_cfg.yaml (even to infinity is fine ;). i.e. you can train a network with only 1 mouse and then it will still predict 2, 3,… 5, 10 - whatever is there.

That makes sense-- thanks very much!

1 Like