search

michalfaber / tensorflow_Realtime_Multi-Person_Pose_Estimation

Multi-Person Pose Estimation project for Tensorflow 2.0 with a small and fast model based on MobilenetV3
Other
216 stars 65 forks source link

Huge difference between Stage branch 1 accuracy and branch 2 accuracy #8

Open dreamstorm25 opened 4 years ago

dreamstorm25 commented 4 years ago

I am training a model based on your repository but it does pose estimation on 13 keypoints. The code has been modified to adapt to new configuration. During training I noticed that there is a huge difference between the L1_acc and L2_acc of any stage. While L2_acc has gone over 0.99 and is steadily increasing, L1_acc is unstable as it increases and decreases without any pattern. L1_acc is about the PAFs and L2_acc is for confidence maps. Could you help me figure out what the issues can occur while predicting PAFs ? diff

Bardhitoo commented 4 years ago

Having the same issue - did you solve it?

dreamstorm25 commented 4 years ago

I have noticed that my results are fine(not always spot on but mostly good), so I am guessing accuracy doesn't play that much of a part in PAFs, it's the loss you should be looking at.

kuanyshbakytuly commented 1 year ago

I am training a model based on your repository but it does pose estimation on 13 keypoints. The code has been modified to adapt to new configuration. During training I noticed that there is a huge difference between the L1_acc and L2_acc of any stage. While L2_acc has gone over 0.99 and is steadily increasing, L1_acc is unstable as it increases and decreases without any pattern. L1_acc is about the PAFs and L2_acc is for confidence maps. Could you help me figure out what the issues can occur while predicting PAFs ? diff

Could you share the code for evaluating accuracy?