GSoC 2020 Application

[Varun Agrawal]

Hi everyone,

I am sorry I'm a little late to the party, but from what I am seeing on the Google Group, my projects of interest haven't been tackled yet.

A little about me: I'm a 3rd year PhD student at Georgia Tech working on robotics research and I heavily use OpenCV in my work. I have been an open source contributor for multiple projects such as ROS, Pytorch (I was responsible for ROI-Pooling and ROI-Align), and even OpenCV (https://github.com/opencv/opencv/pull/9102) but I'd now like to be more heavily involved in contributing and as well as better engage with the community.

My projects of interest are 4, and my final proposal may involve 2-3 (or maybe all 4!) of them, based on mentor feedback on doability.

1. Improve RANSAC in OpenCV: This is very important since a lot of algorithms use RANSAC to eliminate outliers. Additionally, I would like to add support for QDEGSAC  for automatic model selection (https://inf.ethz.ch/personal/marc.pollefeys/pubs/QDEGSAC.pdf)

2. Better SIFT in the main repository: Finally our days of optionally compiling SIFT are over! SIFT is still one of the best feature descriptors out there in its combination of speed, accuracy and light-weightedness. Adding this to the main repo would be a major step forward.

3. Write a tutorial about training a network in TF/PyTorch and then running it in OpenCV: Deep learning is truly upon us and keeping OpenCV has embraced it really well, but this requires more than just code. Given my experience with PyTorch (and deep learning in general) and strong English language skills, writing a tutorial would seem like a great way to contribute to the community. You can sample my writings from my blog http://varunagrawal.github.io/blog/. My older blog is https://computercalledvarun.wordpress.com/.

4. Write a tutorial about using universal intrinsics and cv::parallel_for_ for efficient cross-platform algorithm implementation: Vectorization is hard and honestly, I wasn't aware that OpenCV had such great support for automatic vectorization until I saw this proposal. This seems super useful, though it would involve a fair bit of learning.

I have already written up a partial proposal, but I would love to get feedback from the mentors involved on the priority and scope of the various projects in the proposal. I know I am being a bit optimistic, so any tempering is much appreciated.

Thank you for your time!

[Varun Agrawal]

I would also like to add, I was a GSoC mentor with CloudCV in 2018, so I understand the GSoC process pretty well (if that helps with any potential headaches).