We train dynamic neural radiance fields to predict eventstreams from novel viewpoints and timestamps of a given scene. We share the code for data generation of simulated and real data, as well as code to train and test EvDNeRF models. This codebase was used for the paper EvDNeRF: Reconstructing Event Data with Dynamic Neural Radiance Fields.
First, clone the package and make a couple necessary directories:
git clone git@github.com:anish-bhattacharya/EvDNeRF.git
cd EvDNeRF
mkdir data logs
We provide a docker container that can be pulled from dockerhub.
docker pull evdnerf/evdnerf:dev-user
docker run -d --rm -it --gpus all -v /path/to/EvDNeRF:/home/user/EvDNeRF evdnerf/evdnerf:dev-user
This will run a docker container image, the name of which can be seen in the output of docker ps. Start a shell in the docker container via:
docker exec -it -u root <running_docker_image_name> bash
Now you should be able to run an example train or test configuration within this docker container (see following sections).
(Similar to D-NeRF installation). From inside the EvDNeRF directory:
conda create -n evdnerf python=3.6
conda activate evdnerf
pip install -r requirements.txt
cd torchsearchsorted
pip install .
cd ..
You will have to install additional dependencies manually; we have not provided an updated requirements.txt yet.
We provide datasets public_datasets at the following link to train EvDNeRF models from simulated and real datasets. After extracting the datasets tar, please read the contained, short README.
To train from the jet-down-32 dataset, we will use the configs/jet_train.txt config file. Edit the root_path and datadir to point to your EvDNeRF/ directory and inside the public_datasets directory, respectively.
You may add a ft_file line (refer to the test config file for an example) if you wish to continue training from a weights checkpoint. You may also change the lines i_print (prints training stats to console every i_print iterations), i_img (runs validation and saves event images ang grayscale images every i_img iterations), testskip (one of how many frames to render when doing validation), i_weights (saves weights every i_weights iterations), N_iter (how many iterations to train). We present the best results on simulated data, for example, by training for 100k iterations on 32 event frames (e.g., jet-down-32 dataset), then 50k on 64 (e.g., jet-down-64 dataset), and 50k on 128 (e.g., jet-down-128 dataset). 100k on 32 and 100k on 128 also yields similar results.
Run the following command to start training:
python run_evdnerf.py --config configs/jet_train.txt
This command will run training on the jet-* dataset. Progressive weights and validation are saved to a log folder: ./logs/<exp_datetime>. Any weights file absolute path can then be inputted as the ft_file argument to test on.
Loss and other training statistics are logged with tensorboard. You can view all logged runs in tensorboard by running the following from inside EvDNeRF/, then opening http://localhost:8000/ in your browser.
tensorboard --logdir logs --port 8000
First download pretrained weights. They should placed in the directory EvDNeRF/pretrained_weights. This is the same link as the one above.
There are various testing configurations, including event frame generation directly from json files specifying poses and timestamps (via setting render_test_path), or calculating and comparing metrics against an established test set (via setting render_test).
For an example, we set render_test_path to render predicted events for any defined json file. Various such files are found in test_configs/. test_configs/transforms_test_validation.json contains validation viewpoints each with 32 timesteps. To test a model trained on the jet-down dataset, we will use the configs/jet_test.txt config file, where ft_file points to the corresponding pretrained weights file. Edit the root_path and datadir in the config file to point to your EvDNeRF/ directory and inside the public_datasets directory, respectively.
To test the model, run:
python run_evdnerf.py --config configs/jet_test.txt
When finished, resulting event batches in the form of images are found in ./logs/<exp_datetime>/<exp_datetime>/evim/*.png.
Coming soon for simulated and real data generation!
This codebase is based on the implementation of D-NeRF.
@inproceedings{bhattacharya2024evdnerf,
title={Evdnerf: Reconstructing event data with dynamic neural radiance fields},
author={Bhattacharya, Anish and Madaan, Ratnesh and Cladera, Fernando and Vemprala, Sai and Bonatti, Rogerio and Daniilidis, Kostas and Kapoor, Ashish and Kumar, Vijay and Matni, Nikolai and Gupta, Jayesh K},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={5846--5855},
year={2024}
}