DISTORT-AND-RECOVER-CVPR18

Code for "Distort-and-Recover: Color Enhancement with Deep Reinforcement Learning", CVPR18

Overview

• You can pull&use the docker image from pjc0309/default_setup:latest to run the code. (TensorFlow 0.11.0rc, and other packages such as numpy/scipy)
• Before training,prepare MIT5K train/test images in separate folders (train/raw/, train/target/, test/raw/, test/target/). And edit the path in main.py accordingly.
• run_train.sh starts training.
• Use parse_test.py to parse the test results. (edit the paths accordingly)
• The training speed (iterations per second) should be between 20~40 it/sec. (When trained on i5-6600 and GTX 1080)

Data

• MIT5K Train/Val(RANDOM250) images. Resized to maximum side 500px, JPEG format. (including RANDOM250 list) LINK