Awesome

ColorRL

Tuan Anh Tran, Nguyen Tuan Khoa, Tran Minh Quan, Won-Ki Jeong <br /> ColorRL: Reinforced Coloring for End-to-End Instance Segmentation <br /> In Computer Vision and Pattern Recognition Conference (CVPR), 2021

Paper

Installation

We tested our code with CUDA 10.0, pytorch 1.1.0, gym 0.14.0 for more information about the dependencies, please view dependencies.txt

You can also use docker as follows:

docker pull anhtuanhsgs/pytorch-openai:1.1

Data Preparation

For 3D datasets, we use Tag Image File Format (TIFF) format. For 2D images, we tested our code with both .png and .tif files. Input images and their labels should be placed in two folders: A and B, respectively. For example:

path_to_train_set/A/*.tif (for input images)
path_to_train_set/B/*.tif (for label images)

Testing data path is setup as:

path_to_test_set/A/*.tif

To update data path, please modify main.py

Evaluation

For evaluation details, visit jupyter notebooks in evaluation/

Example with CREMI

We include our Cre-256 dataset in '''Data/Cremi/Corrected''' For parameters' usage, please see '''main.py'''

To train ColorRL agent: (adjust the number of GPUs and number of workers that are best for your system before running the script)

bash run_scrips/256_cremi_train.sh

To deploy a trained agent: (need to modify the checkpoint path to a saved checkpoint)

bash run_scrips/256_cremi_deploy.sh

Step-by-step example with CVPPP

Setting up for CVPPP data set can be done as follows:

Download CVPPP data from https://www.plant-phenotyping.org/CVPPP2017 and extract the .h5 files to '''Data/CVPPP_Challenge/''' then run

mkdir -p Data/CVPPP_Challenge/train/A/
mkdir -p Data/CVPPP_Challenge/train/B/
mkdir -p Data/CVPPP_Challenge/valid/A/
mkdir -p Data/CVPPP_Challenge/valid/B/
mkdir -p Data/CVPPP_Challenge/test/B/
cd Data/CVPPP_Challenge/
python ExtractData.py

Training

For training with CVPPP (similarly with other data), run:

bash run_scrips/cvppp_train.sh

Tensorboard can be used for training logs, use:

tensorboard --logdir=logs/

checkpoints are saved at trained_models

Inference

To make predictions on a dataset, For test set inference with CVPPP (similarly with other data), edit run_scrips/cvppp_deploy.sh:

--load: path to a check point (eg. trained_models/cvppp/cvppp/). Path to a checkpoint will have the following format: [save_model_dir]/[data]/[env]_[model]/xxxx.dat

for example: with:

--env 256_cremi_train
--model AttUNet2
--data 256_cremi
--save-model-dir trained_models/

then an example of loading a checkpoint would be: --load trained_models/256_cremi/256_cremi_train_AttUNet2/15000.dat

--deploy: to run as an inference task

To make predictions, run:

bash run_scrips/cvppp_deploy.sh

Results are stored at deploy/