GSoC Interest – Contributing to DeepForest

[Bhavya1604]

Dear DeepForest Community,

I hope you're doing well! While exploring GSoC project ideas, I came across two projects listed under weecology/retriever that caught my interest. Since both projects mention DeepForest as the source code, I wanted to dive deeper into this repository and understand how I can contribute effectively.

1. Efficient Detection of Unique Images from Overlapping Images – This project focuses on removing double counting in overlapping images using either the weecology/DoubleCounting or open-forest-observatory/geographer repository. I’d love to understand which repository would be the best starting point and any key areas of the DeepForest codebase I should explore.

2. Developing an Active Learning Module for DeepForest – I’m particularly excited about this because of its potential to improve model training efficiency. I’d like to integrate the BOEM repository’s active learning code into DeepForest and develop a user-friendly image selection interface. Any suggestions on how to approach this would be great.

I’d love to hear your thoughts on:

• How I should get started with each project.
• Which parts of the DeepForest codebase I should focus on.
• Any past contributions or examples that could help.

Looking forward to your insights and excited to contribute!

Best regards,
Bhavya Mehta

[Bhavya1604]

Hi @bw4sz , @henrykironde , @ethanwhite , I’d love to hear your thoughts on this when you get a chance. Appreciate any guidance!

[henrykironde]

Thank you for your interest in the project! To get started, we recommend reading through the documentation, particularly the developer section. This will help you gain the experience needed to showcase your skills by tackling some issues. We've labeled certain issues as "good first issues" to provide additional insight into the design and structure of the codebase.

[Bhavya1604]

Thank you @henrykironde ,
Thanks for the response! I’ve gone through the docs, set up DeepForest in a virtual environment, ran the tests, and tested the model on an image so I’ve got a basic understanding now. While I’ll continue exploring the codebase and related repos on my own, it would really help if you could provide some guidance on where to start with the "Efficient Detection of Unique Images from Overlapping Images" project.

[henrykironde]

@Bhavya1604 let us break down the contents in the approach:

Choose a suitable repository either weecology/DoubleCounting or open-forest-observatory/geograypher for implementing the unique image detections workflow.
Develop an efficient algorithm for removing double counting among overlapping images.
Evaluate the performance of the workflow on various datasets.

I recommend starting by choosing the repository you’d like to work with either, weecology/DoubleCounting or open-forest-observatory/geograypher. Take time to understand the repo’s workflow, as they provide sample datasets to explore and experiment with. This hands-on approach will give you a clearer understanding of how to apply solutions to the DeepForest workflow effectively.

[Bhavya1604]

I’ve set up the weecology/DoubleCounting repo, installed dependencies in a virtual environment, and ran pytest. However, I’m running into multiple test failures. Some errors seem to be related to missing attributes in certain functions, while others mention missing files or dependencies.
I’m also noticing that some tests generate blank outputs—figures with just X and Y axes but no actual content.
could you guide me on what might be causing these issues?

[bw4sz]

This is interesting, pycolmap and these stats are poorly documented and move quickly. In general we will be trying to lift code from this repo, so best to just a get a sense for what it does and the component pieces. Also check out https://github.com/open-forest-observatory/geograypher/blob/main/examples/project_detections.ipynb

[Bhavya1604]

Thank you for the response! I'll explore the repo to understand how the code works and take ideas.
I'll also go through the geograypher example to get a better sense of how it works.

[Bhavya1604]

@bw4sz I have found the solution to these above errors in the DoubleCounting repo, most of it is new names for functions in pycolmap lib and few more things, should i address these in a PR there? so that new users don't face them.
you can see it here #1098 that i have used the new function estimate_homography_matrix that replaces the old homography_matrix_estimation
the repo has a clean record (like no issues history) so i wanted to confirm this with you

[Bhavya1604]

Hi @henrykironde, I’ve started writing a proposal for this project. How should I share it with you to get your feedback?

[Bhavya1604]

Hi @bw4sz @henrykironde @ethanwhite, I have gone through both the repository and the reference research papers in them and after all that I am proposing a method to solve the double counting problem in overlapping images. The idea is to combine three different techniques to make sure we don't count the same tree or animal twice when it appears in multiple images. First, we create what I call a "confidence map" think of it like a heat map showing how sure we are about each detected object. Then, we look at how objects in different images overlap spatially - kind of like checking if two circles overlap on a map. Finally, for any cases we are not sure about, we compare the actual features of the objects (like comparing two photos to see if they show the same tree) using deep learning. What makes this approach better is that it's like having multiple checks in place so if one method isn't sure, the others can help make the decision. It's also efficient because we only do the complex feature comparison when we really need to, and it fits nicely with how DeepForest already works

let me know what you think about this idea

[bw4sz]

I'm fine with extending the idea after first showing the previous implemented ideas are not sufficient. GSOC proposals are much more about inserting functionality into open-source projects than novel research. I'd need to see clear examples about how the ideas in the DoubleCounting repo and the geograypher package are not sufficient. The bar would be pretty high given the difficulty. If this doesn't suit your interests, I would suggest you look at different projects or proposals.

[Bhavya1604]

Thanks for clarifying that the primary focus is on integrating the existing repositories and that new ideas are optional.

This project is very interesting to me and while the bar is high, I see that as a challenge rather than a barrier. In fact, the difficulty makes me more dedicated to solving it rather than giving up, which is why I believe this is the perfect project for me.

I have reviewed both repositories, and I have some doubts about certain aspects of the project. Would it be possible to set up a meeting to discuss them? That way, I can get clarity and focus on the right direction moving forward.