Authors: Hajin Choi, Bochang Moon

This is an official implementation of the paper "Consistent Post-Reconstruction for Progressive Photon Mapping", which is presented in Pacific Graphics 2021. This repository includes training/test codes with four test images needed to reproduce the results in the paper. For re-training using other images, please refer to the paper and scripts/config.py to see how the input images should be consist of. We recommend the Mitsuba renderer for generating the SPPM images.

• OS: Ubuntu 20.04
• CPU: AMD Ryzen Threadripper 3990X
• GPU: NVIDIA RTX 3090

• NVIDIA graphic driver (tested by version >= 460)
• Docker
• nvidia-docker
  • https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker

1. Clone this repository
2. Inside the cloned directory, build a docker image by
   docker build . -t IMAGE_NAME
3. Create a container by
   docker run -d -it -v $PWD:/consistent-recon --gpus all IMAGE_NAME

• For training, set following things in scripts/config.py

  # Mode select (TRAINING, TEST)
  config["mode"] = "TRAINING"
  # Number of epochs
  config['epochs'] = 100
  # Prefix for dataset directory. Note that it will be used to find the refereces too (e.g., dataset_refereces). Same for test.
  config["datasetPrefix"] = "dataset"
  # Directory of training dataset
  config["trainDatasetDirectory"] = "dataset_train"
  # List of scenes (e.g., dataset_train/box)
  config["trainScenes"] = ['box', 'sponza']

• For test, set following things in scripts/config.py

  # Mode select (TRAINING, TEST)
  config["mode"] = "TEST"
  # Epoch of checkpoint to load
  config["loadEpoch"] = "100"
  # Directory of test dataset
  config["testDatasetDirectory"] = "dataset_test"
  # List of scenes (e.g., dataset_test/box)
  config["testScenes"] = ['bookshelf', 'breakfast-room', 'pool', 'water']

• Others are optional and can be left at their default values.
• Outputs of the test will be stored in /consistent-recon/output

1. Attach to the container by a docker command e.g.,
   docker exec -w /consistent-recon -it CONTAINER_ID bash
2. Run the code by python scripts/main.py
   • If you meet a weird memory increase along the epochs, then run the script with TCMalloc
LD_PRELOAD="/usr/lib/x86_64-linux-gnu/libtcmalloc_minimal.so.4.3.0" python scripts/main.py

All source codes are released under a BSD License.

We used exr-related operations using the script scripts/exr.py from KPCN.