GSoC 2025 Proposal 1: Efficient Detection of Unique Images from Overlapping Images

[gadhvirushiraj]

Hello @bw4sz, @ethanwhite, and @henrykironde,

My name is Rushiraj Gadhvi, a fourth-year Computer Science student majoring in AI at Plaksha University. I discovered the DeepForest library a few months ago while researching hackathon ideas and was immediately impressed by the innovative work being done.

I recently reviewed this year’s GSoC project proposals, and both the "Efficient Detection of Unique Images from Overlapping Images" and the "Active Learning Project" captured my interest. The challenges they present align perfectly with my academic and research interests, and I am eager to contribute and expand my skills through GSoC.

At present, I am delving into Project 1. I have gone through all the papers referenced in the Double-Counting repository, I began ideating on potential approaches and have a couple of questions:

• I tested DFM on my M1 Air (8GB), but it proved quite slow. Could you please share your thoughts on the balance between processing time and accuracy in this project? @bw4sz . I’m considering experimenting with SIFT or AKAZE next to evaluate their performance.

• Additionally, is there any specific reason for recommending deep learning extraction methods over traditional methods like SIFT? (I presume beacuse it is hard to find unique feature in dense vegetation background). I would also appreciate knowing if any other methods have been tried previously for solving the double counting issue besides those mentioned in the Double-Counting repository.

Thank you.

[henrykironde]

Related to #991 authoredBy: @Karanveer266

I noticed that the current implementation in the DoubleCounting repository uses iterative homography matching for image alignment. While effective, this approach can be computationally intensive, potentially limiting its scalability when processing large datasets. Given that one of the project goals is to evaluate the performance on various datasets, this could present a significant challenge.

I would like to inquire:

Are we expected to adhere to the current iterative homography matching approach?

If not, would you be open to exploring alternative methods that could potentially improve processing efficiency, especially for larger datasets?

Do you have any specific performance targets or dataset sizes in mind that we should be aiming for?

I am eager to discuss potential optimization strategies and ensure that our approach aligns with the project's goals.

[bw4sz]

See my answer here

#953 (reply in thread)

We have several datasets and procure more. We always prioritize simple existing solutions over complex novel ones. The first step is to demonstrate that a simple solution fails in most cases before we move to anything more sophisticated, that is how you get buy-in from a community and avoid wasting engineering time.