search

spoonsso / dannce

MIT License
214 stars 30 forks source link

Multi animal COM #140

Closed histun closed 1 year ago

histun commented 1 year ago

Hi

I was able to get a single mouse COM/DANNCE tracking fine, but having a difficulty with two instances. Initially, I thought maybe because two mice look similar (only one has an implant), so I tested a mouse with a food object (shown below). Often, predicted results will give 2 COMS overlaid or/and in between. (two nearby frames) capture capture1

Below is from the train_results/debug_im_out_valid. I see separate images with 1 COM (correct position) only. Is this expected? capture3

I'm finetuning the weights.rat.COM.hdf5 just like I did for single animal COM training. Only thing I changed in the config was adding n_instance: 2. I'm wondering if I'm missing anything.

histun commented 1 year ago

Training with more labels from scratch helped solve the issue. Thanks!

data-hound commented 1 year ago

Closing this since it is not an issue anymore. Please reopen if you need help.