Video reconstruction by spatio-temporal fusion of blurred-coded image pair

This repository contains inference code for Video reconstruction by spatio-temporal fusion of blurred-coded image pair accepted at ICPR 2020.

Preprint version of the paper is available here

Dependencies

python v3.6.8
pytorch v1.4.0
numpy v1.16.4
skimage v0.15.0

Input images and model weights

data/test_videos: contains 14 test video sequences of 9 frames each, randomly selected from GoPro dataset and used for evaluation; input to the network is obtained by coded exposure of these frames
weights: Download the model weights to this directory from the following paths:

Weights for coded-blurred input: download model
Weights for single coded input: download model
Copy the downloaded .tar.gz files to the weights directory and extract using

tar -xvzf single-coded-inp.tar.gz
tar -xzvf coded-blurred-inp-attn.tar.gz

Video reconstruction from coded-blurred image pair

Command to run inference on test videos in data/test_videos and optionally save results in recon_results:

python recon_cb.py --savepath recon_results

Video reconstruction from single coded exposure image

Command to run inference on test videos in data/test_videos and optionally save results in recon_results:

python recon_sc.py --savepath recon_results

Supplementary material

Video reconstruction results from the paper can be viewed in the supplementary material here

Bibtex

@article{shedligeri2020video,
  title={Video Reconstruction by Spatio-Temporal Fusion of Blurred-Coded Image Pair},
  author={Shedligeri, Prasan and Pal, Abhishek and Mitra, Kaushik and others},
  journal={arXiv preprint arXiv:2010.10052},
  year={2020}
}