search

GeorgeSeif / Semantic-Segmentation-Suite

Semantic Segmentation Suite in TensorFlow. Implement, train, and test new Semantic Segmentation models easily!
2.5k stars 880 forks source link

gt pictures in checkpoints different with val_labels? #202

Closed x7hkvip closed 5 years ago

x7hkvip commented 5 years ago

what I have done: Papered my own dataset. I wrote a script to convert 8 bit gray label pictures into 24 bit RGB pictures. Then I modified csv files. what's the problems? When training,I found the ground truth pictures in checkpoints files are different to the lables in val_labels files(quite different!). The predit pictures I got are all totally black pictures,but the current_loss isn't equals to 0. What should I do?Very appricate for your help!

x7hkvip commented 5 years ago

I fixed it.The reason why I got this problem because when after I set color to my 8 bit gray picture,I convert it from png to jpg,this step add a lot of strange RGB values into my pictures.

zihao0821 commented 5 years ago

I fixed it.The reason why I got this problem because when after I set color to my 8 bit gray picture,I convert it from png to jpg,this step add a lot of strange RGB values into my pictures.

Hello,

I use opencv to covert a .png to .jpg, and face the same problem. Can you tell to deal with that?

thanks！

zihao0821 commented 5 years ago

I fixed it.The reason why I got this problem because when after I set color to my 8 bit gray picture,I convert it from png to jpg,this step add a lot of strange RGB values into my pictures.

Oh, I find a way. Just convert a 8 bit .png to 24 bit .png :)