search

michalfaber / keras_Realtime_Multi-Person_Pose_Estimation

Keras version of Realtime Multi-Person Pose Estimation project
Other
780 stars 372 forks source link

Masked the outputs #124

Open rafikg opened 5 years ago

rafikg commented 5 years ago

Thanks @michalfaber, I am wondering if still make sense to mask the outputs if we have already masks the input image before. If we remove the outputs mask, we don't still need two build two different graphs for training and testing?

michalfaber commented 5 years ago

I experimented a little bit with masking the input image instead of the masking outputs. I didn't see a significant improvement so I left the model unchanged. The original paper says that masking improves the performance by 2.3 % but they were training their model much longer.

rafikg commented 5 years ago

@DeeperDeeper, Thanks, I definitely agree!