search

yenchenlin / pix2pix-tensorflow

TensorFlow implementation of "Image-to-Image Translation Using Conditional Adversarial Networks".
MIT License
939 stars 300 forks source link

Training day-night #19

Open rudolphyo opened 7 years ago

rudolphyo commented 7 years ago

Had anyone tried to reproduce results of day to night?

The problem is that i get very poor results. I take dataset mentioned in paper and make folders A and B in this way : take strong day photos(daylight > 0.8) for A and strong night photos (night > 0.6) for B for each webcam. Then i made some data augmentation through rotating/mirroring of images. After that I combined them with script and change batch_size from 1 to 4 and epochs from 200 to 17 as in paper.

Thank you!

ppwwyyxx commented 7 years ago

Did you make sure to apply identical augmentations to each training/testing pair?

rudolphyo commented 7 years ago

@ppwwyyxx yeah, sure. If in A i have 90 left rotated image, so in B it is the same, but night one. I can share my script that I`ve used for augmentation

AshuKV commented 5 years ago

@rudolphyo Can you please share your github link or your training and testing scripts for your datasets ? And how did you train it?

fabiosoft commented 4 years ago

I get a very blurry and weird artifacts in results. --batch_size 4 --preprocess scale_width_and_crop --load_size 754 --crop_size 512