Qifeng Chen

@yaxingwang You need 12Gb GPU memory to run the code.

If you just the weights listed there, it should be fine. I use tf.reduce_mean to compute the average. The weights 1,1/1.6,1/2.3,1/2.8,10/0.5 are computed from the average l0,...,l5 in the 100...

1. That's correct. 2. I expand a 19D label map to 20D. The expanded extra dimension is for "void" regions. This ensures that each pixel has a one-hot-vector at each...

You can obtain the data from https://download.visinf.tu-darmstadt.de/data/from_games/

For the users in China, I share the Tensorflow models in Baidu Cloud: http://pan.baidu.com/s/1dEWOFU5

You can download the models from here: https://drive.google.com/open?id=0B_B_FOgPxgFLRmJVb1VYbEJWR0k I will update the GitHub with direct URLs.

You can download the model manually from my personal website: https://cqf.io/ImageSynthesis/Models/

I just update the code in a folder "extract_cityscapes_input_data". You can use it to extract input labels and color images.

You also need to download the original Cityscapes data from https://www.cityscapes-dataset.com. You may need to go through "extract256full.m" to understand the process.

In the VGG network, these three values are the mean RGB values (in the ImageNet) to normalize the input image