Hi, I'm using SLEAP to predict single animal poses. I'm importing two videos of one single fruit file in grayscale, recorded at 100fps, resolution of 900 x 900 and with ~ 10000 frames each. I chose and labeled 10 random frames of each video and started to train my model using basic hyperparameters. So far, I've ran over 10 training sessions and I'm facing one of the three situations every time I resume training: 1) no instances are predicted in the frames, 2) instances are predicted in all frames and metrics are not generated or 3) instances are predict, metrics are generated and they depict a real poor model performance (I've had situations where OKS mAP = 0), regardless of correcting of previous predictions. I've adjust the hyperparameters several times and, more recently, I've tried to use the training configurations of a previous model from other experiment that showed good metrics to run my training now. Nothing is working out.

I've been a SLEAP user for a while now and I've successfully managed to train a model in the past using very identical settings to the ones I'm using now (e.g., video settings of the images I'm importing to my project and training settings (i.e., skeleton, labeling, model hyperparameters)). Am I missing something? I've reached almost 300 labeled frames and I see no improvement at all...

Would be very grateful for you help!

Thanks,

Rita