Poor performance issue

What is the typical performance for mice pose estimation (single C57 w/ fiber implant on head, in home cage with bedding, marker at nose, l&r ear, tailbase)? I found the nose and the tailbase mark having poor performance.
I tried resnet50 and mobilenet2, with iteration number 100000 and 150000.

Any suggestions? Thanks very much!

1 Like

depends on (1) how many labels you made; (2) how good your labeling is.

Please see the published papers for many discussions and experiments on this: www.deeplabcut.org

Thanks for the reply. I labeled 20 frames as default, and for each one I labeled nose, leftear, rightear and tailbase. By eye I believe the labeling should be accurate.
I have read the Nat Neurosci and the Nat Protoc paper, and I noticed ā€œas few as 50 frames could be labeled for accurate tracking of a mouse snout in an open-ļ¬eld-like scenario with challenging background and lighting statisticsā€. So is 50 a typical number for manual labeling? Thanks!

20 is a random default, and that is per video; the dataset should be from 5-10 different videos of a behavior you wish to track. Note we labeled 1,080 frames in the NN paper for benchmarking to determine that ~50 frames was <10 pixel error - which was biologically meaningful. We also recommend training for at minimum 200K iterationsā€¦


Oh I see, thanks for the suggestions! Then may I ask how to put 5-10 videos into one training dataset? Thanks!

when you create the project you can add a whole folder of videos; if already created, you use the function add_new_videos then you can label 10-20 frames per video, and that would be better performanceā€¦

Thanks for your suggestions. It works now!