DeepLandscape: Adversarial Modeling of Landscape Videos

Implementation of DeepLandscape: Adversarial Modeling of Landscape Video in PyTorch

Project Page | Video Explanation | Paper | Teaser [1min]

Official repository for the paper E. Logacheva, R. Suvorov, O. Khomenko, A. Mashikhin, and V. Lempitsky. "DeepLandscape: Adversarial Modeling of Landscape Videos" In 2020 European Conference on Computer Vision (ECCV).

GIF sample

Requirements

pip3 install -r requirements.txt

Usage

Download everything from here and put in the results directory.

Homographies

Use homographies/manual_homographies to reproduce the paper; use homographies/manual_homographies_x2.5 if you want the speed to match the speed of real videos in test data; use homographies/selected_homographies to get the best visual results.

Animate Images

Move images you like to animate to results/test_images. Then run

PYTHONPATH=`pwd`:$PYTHONPATH runfiles/encode_and_animate_test_all256_with_style.sh <homography_dir>

Results will be saved in results/encode_and_animate_results.

Generate Synthetic Videos

To use the 256x256 generator run

./generate.py config/train/256.yaml --homography_dir <homography_dir>

Results will be saved in results/generated

Calculate Metrics

TBD

Super-Resoultion

See superres/README.md

Train

Train the Generator

You should prepare an lmdb dataset:

./prepare_data.py <data type (images or videos)> <input data path> --out <output lmdb directory> --n_worker <number of workers>

To train the 256x256 generator:

./train.py configs/train/256.yaml --restart -i <path to image data> -v <path to video data>

Train the Encoder

You should prepare a dataset:

PYTHONPATH=`pwd`:$PYTHONPATH runfiles/gen_encoder_train_data_256.sh

To train the 256x256 encoder:

PYTHONPATH=`pwd`:$PYTHONPATH runfiles/train_encoder_256.sh

Download Test Data

vid_dl/main.py <output directory>

Acknowledgment

This repository is based on Kim Seonghyeon's Implementation A Style-Based Generator Architecture for Generative Adversarial Networks in PyTorch

The Superresolution part is based on https://github.com/xinntao/BasicSR

Mean optical flow calculation is taken from https://github.com/avinashpaliwal/Super-SloMo

Segmentation is taken form https://github.com/CSAILVision/semantic-segmentation-pytorch

Metrics

LPIPS metric https://github.com/richzhang/PerceptualSimilarity
SSIM https://github.com/Po-Hsun-Su/pytorch-ssim
FID https://github.com/mseitzer/pytorch-fid

Citation

If you found our work useful, please don't forget to cite

@inproceedings{Logacheva_2020_ECCV,
  author = {Logacheva, Elizaveta and
            Suvorov, Roman and
            Khomenko, Oleg and
            Mashikhin, Anton and
            Lempitsky, Victor
  },
  title = {DeepLandscape: Adversarial Modeling of Landscape Videos},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  month = {August},
  year = {2020},
}

advimman/deep-landscape