search

Engineering-Course / LIP_JPPNet

Code repository for Joint Body Parsing & Pose Estimation Network, T-PAMI 2018
MIT License
341 stars 87 forks source link

Could I get the better pre-training model? The performance of the pre-training model is slightly worse than that described in the paper. #48

Open taoshiqian opened 5 years ago

taoshiqian commented 5 years ago

The performance of the pre-training model is slightly worse than that described in the paper, especially the Mean Accuracy.

They are the results of using the pre-trained model on the validation set.

  1. Am I doing something wrong? Evaluation code reference link: https://github.com/HCPLab-SYSU/ATEN/blob/master/evaluate/test_parsing.py
  2. Is there a better pre-training model?

==================================================

overall accuracy 0.8625252479197769

Accuracy for each class (pixel accuracy): background : 0.935942 hat : 0.784895 hair : 0.801849 sun-glasses : 0.416315 upper-clothes : 0.240663 dress : 0.825752 coat : 0.344182 socks : 0.680521 pants : 0.513359 gloves : 0.847266 scarf : 0.322934 skirt : 0.158371 torso-skin : 0.329161 face : 0.858112 right-arm : 0.730865 left-arm : 0.752589 right-leg : 0.695891 left-leg : 0.708378 right-shoe : 0.583139 left-shoe : 0.589186 mean accuracy 0.6059685085404918

background : 0.862065 hat : 0.635582 hair : 0.693225 sun-glasses : 0.340043 upper-clothes : 0.212232 dress : 0.678585 coat : 0.289160 socks : 0.548446 pants : 0.421071 gloves : 0.717109 scarf : 0.250269 skirt : 0.142373 torso-skin : 0.238828 face : 0.730600 right-arm : 0.619979 left-arm : 0.634775 right-leg : 0.572062 left-leg : 0.571267 right-shoe : 0.441749 left-shoe : 0.446395 mean IU 0.5022908042317561

fwavacc 0.7652094827965213