A Unified Framework for Compressive Video Recovery from Coded Exposure Techniques

This repository contains the official implementation of the work A Unified Framework for Compressive Video Recovery from Coded Exposure Techniques accepted to be published at IEEE/CVF WACV 2021.

args used in evaluating the model

• --ckpt : refers to the name of the model to be used
• --save_gif : saves the ground truth and predicted frames to the disk. Otherwise the code only logs the PSNR values.
• --flutter : to be used when evaluating the flutter shutter model
• --two_bucket : to be used when evaluating the two bucked coded-blurred image pair model

Evaluation

We provide evaluation code for three different and important models in the paper. We provide the DNN test set (described in the paper) for evaluation, in the data directory.

• Flutter shutter for 16x reconstruction : download model
• Pixel wise coded exposure for 16x reconstruction : download model
• Coded-Blurred pair for 16x reconstruction: download model

Download the appropriate model file from the links provided above and copy them to the models directory. The files are compressed in .tar.xz format which can be extracted by tar -xvf <filename>.

1. Evaluating flutter shutter model:

python infer_h5.py --ckpt flutter_optimal.pth --gpu 0 --save_gif --flutter

2. Evaluating the pixel-wise exposure model:

python infer_h5.py --ckpt pixel_optimal.pth --gpu 0 --save_gif

3. Evaluating the pixel-wise exposure model:

python infer_h5.py --ckpt c2b_optimal.pth --gpu 0 --save_gif --two_bucket