search

idearibosome / tf-perceptual-eusr

A TensorFlow-based image super-resolution model considering both quantitative and perceptual quality
Apache License 2.0
45 stars 10 forks source link

What is the dataset you have used... #1

Closed vdsprakash closed 5 years ago

idearibosome commented 5 years ago

We trained our model with the DIV2K dataset, which was used in the NTIRE challenge. https://data.vision.ee.ethz.ch/cvl/DIV2K/ Our model supports multiple upscaling factors in the same model (e.g., x2, x4, x8), however, the original dataset does not provide training images downscaled by a factor of 8. Therefore, you have to manually downscale the high resolution (HR) images.

idearibosome commented 5 years ago

Updates: We added a helper script to obtain downscaled images (train/misc/downscale_generator_matlab.m). Just download the high resolution (HR) images from the official DIV2K dataset website and generate downscaled images by the provided script. Detailed instruction is also provided in train/README.md.