Novel View Synthesis via Depth-guided Skip Connections

Code for paper Novel View Synthesis via Depth-guided Skip Connections

@InProceedings{Hou_2021_WACV,
    author    = {Hou, Yuxin and Solin, Arno and Kannala, Juho},
    title     = {Novel View Synthesis via Depth-Guided Skip Connections},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {January},
    year      = {2021},
    pages     = {3119-3128}
}

Data

Download the dataset from the Google Drive provided by [1] and unzip the dataset under ./datasets/ folder.

Download the evaluation list files from the Google Drive. Put the corresponding file under corresponding dataset folder. E.g. ./datasets/dataset_kitti/eval_kitti.txt.

Training

python train.py\
    --name chair\
    --category chair\
    --niter 2000\
    --niter_decay 2000\
    --save_epoch_freq 100\
    --random_elevation\
    --lr 1e-4

If you don't want to view the real-time results, you can add command --display_id 0

If you want to view training results and loss plots, run python -m visdom.server and click the URL http://localhost:8097.

Testing

Download our pre-trained model from our Google Drive To evaluate the performance, run

python eval.py\
    --name chair\
    --category chair\
    --checkpoints_dir checkpoints\ 
    --which_epoch best

Acknowledgments

The code is based on the source code of the paper:

[1] Chen, Xu and Song, Jie and Hilliges, Otmar (2019). Monocular Neural Image-based Rendering with Continuous View Control. In: International Conference on Computer Vision (ICCV). (https://github.com/xuchen-ethz/continuous_view_synthesis),