This repository contains code for our RSS 2020 paper. Official proceedings is available here and pre-print is available at https://arxiv.org/abs/2006.04973.

@INPROCEEDINGS{Ramanagopal-RSS-20, 
    AUTHOR    = {Manikandasriram Srinivasan Ramanagopal AND Zixu Zhang AND Ram Vasudevan AND Matthew Johnson Roberson}, 
    TITLE     = {{Pixel-Wise Motion Deblurring of Thermal Videos}}, 
    BOOKTITLE = {Proceedings of Robotics: Science and Systems}, 
    YEAR      = {2020}, 
    ADDRESS   = {Corvalis, Oregon, USA}, 
    MONTH     = {July}, 
    DOI       = {10.15607/RSS.2020.XVI.022} 
}

We treat motion deblurring as a per pixel temporal problem. We used FLIR A655sc radiometrically calibrated camera to record images at 200Hz i.e. the sampling period (5ms) is roughly half the thermal time constant (~11ms) of the camera. We are providing sample data of an outdoor sequence with the camera panning horizontally as a .mat file here

1. You need IBM cplex optimization studio available here. A free academic version is available for students and researchers. In particular, install the python API for cplex.

2. Other required python packages:

numpy
matplotlib
scikit-image
scipy
h5py
tqdm

• Clone repository and add to PYTHONPATH

git clone https://github.com/umautobots/pixelwise-deblurring.git
cd pixelwise-deblurring
export PYTHONPATH=<path/to/pixelwise-deblurring>:$PYTHONPATH

• To deblur <N> frames starting from <start_num>

python3.6 ./src/deblur_main.py --matfile <path/to/downloaded/data> --indices <start_num> --N <N> --output-prefix <path/to/output/folder>

You can provide comma separated list of starting indices and output files are automatically named as {output_prefix}_{start_num}_{start_num+N}.npz

Note: Since each pixel is independently processed, ~80k optimization problems needs to be solved which is currently slow. The code will automatically use the maximum number of CPU cores available for parallel processing.

• To view the processed files, use:

python3.6 ./src/view_processed_frames.py --filename <path/to/npz/file>