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

After 100 Epoch Total Loss Is Still Very High #22

Open rhiver opened 3 years ago

rhiver commented 3 years ago

The loss is about 2000-3000, the loss doesn't change much from epoch 3 to epoch 100. Anything wrong with it？

michalfaber commented 3 years ago

I've observed a decreasing loss for around 100 epochs but, indeed the value remains too high. I like the idea used in HRNet-Bottom-Up-Pose-Estimation where they apply clever masks before calculating the loss. The masks contain 1s at the joints location, 0s at the crowd or 0-keypoints segmentations, remaining pixels contain background color.