Skip to content

Latest commit

 

History

History
91 lines (73 loc) · 4.58 KB

README.md

File metadata and controls

91 lines (73 loc) · 4.58 KB
Raw

Learning to Render Novel Views from Wide-Baseline Stereo Pairs

Project Page | Paper

Explore in Colab

Yilun Du, Cameron Smith, Ayush Tewari, Vincent Sitzmann
MIT

This is a official implementation of the paper "Learning to Render Novel Views from Wide-Baseline Stereo Pairs".

Google Colab

If you want to experiment with our approach, we have written a Colab. It loads two input images and estimates the underlying fundemental matrix between images. It doesn't require installing anything, and illustrates and renders videos when interpolating between the two given images.

Get started

You can install the packages used in this codebase use the following command

pip install -r requirements.txt

You will need to download the Realestate10k and ACID dataset. We have attached a small subset of the RealEstate10K dataset as well as pose files which we may directly download using the dropbox link here. Please extract the folder in the root directory of the repo.

To download the full datasets for Realestate10k and ACID, please follow the README here.

You can also load a pretrained model on the RealEstate dataset here. Please extract the model in the root directory of the repo.

High-Level structure

The code is organized as follows:

  • ./dataset/ contains code loading data
  • ./data_download/ contains code to download Realestate10k and ACID datasets
  • ./utils/ contains code for different utility functions on the dataset
  • ./estimate_pose/ contains code for estimate the pose between two images using superpoint
  • models.py contains the underlying code for instantiating the model
  • training.py contains the underlying code for training the model
  • loss_functions.py contains loss functions for the different experiments.
  • summaries.py contains summary functions for training
  • ./experiment_scripts/ contains scripts to reproduce experiments in the paper.

Reproducing experiments

The directory experiment_scripts contains one script per experiment in the paper.

To monitor progress, the training code writes tensorboard summaries into a "summaries"" subdirectory in the logging_root. The code will run infinitely -- you may stop the code after a day or two of training.

Realestate experiments

The Realestate experiment can be reproduced by running the command below for two days

python experiment_scripts/train_realestate10k.py --experiment_name realestate --batch_size 12 --gpus 4

followed by running the command below (adding lpips and depth loss later in training)

python experiment_scripts/train_realestate10k.py --experiment_name realestate_lpips_depth --batch_size 4 --gpus 4 --lpips --depth --checkpoint_path logs/realestate/checkpoints/model_current.pth

You can evaluate the results of the trained model using

python experiment_scripts/eval_realestate10k.py --experiment_name vis_realestate --batch_size 12 --gpus 1 --views=2 --checkpoint_path logs/realestate/checkpoints/model_current.pth

You can also visualize the results of applying the trained model on videos using the command

python experiment_scripts/render_realestate10k_traj.py --experiment_name vis_realestate --batch_size 12 --views 2 --gpus 1 --checkpoint_path logs/realestate/checkpoints/model_current.pth

and on unposed images using

python experiment_scripts/render_unposed_traj.py --experiment_name vis_realestate --batch_size 12 --views 2 --gpus 1 --checkpoint_path logs/realestate/checkpoints/model_current.pth

Citation

If you find our work useful in your research, you can cite the following article:

@article{du2023widerender,
          title={Learning to Render Novel Views from
            Wide-Baseline Stereo Pairs},
          author={Du, Yilun and Smith, Cameron and Tewari,
                Ayush and Sitzmann, Vincent},
          journal={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
          year={2023}
}

Contact

If you have any questions, please feel free to email the authors.