Discussion about GSoC projects

[Mu-Magdy]

Hi There,

I'm really keen on your GSoC projects after following your work last year. I'm having trouble choosing one because they're all so interesting, and I believe I have the skills for many of them. Is there a project you're eager to prioritize? Knowing this would help me decide.

[henrykironde]

We don't prioritize projects; our main goal is to find individuals capable of delivering and completing them. Therefore, we typically select at least one contributor per project, with preference given to those consistently delivering substantial contributions. While minor contributions such as fixing typos are appreciated, they usually don't weigh heavily in the selection process. Additionally, contributions don't necessarily have to be merged to be considered substantial.

[Mu-Magdy]

I can modify it if you want to or I can restructure it completely or extend the timeline to add more weeks if needed, But I am afraid to miss the deadline for the proposal submission.

[Mu-Magdy]

Hi @henrykironde,

Is there a modification in the proposal?
Should I submit it or not?

[ethanwhite]

Hi @Mu-Magdy - I have it a quick read and everything looks good! Definitely submit before the deadline even if you don't hear from Henry. Good luck!

[henrykironde]

Looks good to me. Sent you an email

[Mu-Magdy]

Hi @henrykironde,
What do you advice me to do now?
I read the papers @bw4sz mentioned in his comment. And explored the model that he shared.
How to better prepare for this project ?

[bw4sz]

Hello @Mu-Magdy here are some details on the status of the project.

1. We have a large number (1000s) of photos taken from airplanes of wading bird colonies. Here is a low resolution screenshot

2. We have images and a trained model from UAV data. These images are higher resolution and easier to identify. I can share the trained model with you. This UAV model works reasonably well on Airplane imagery. Sample predicted airplane image ->

4. We have annotations from scientists on Airplane imagery that you can use for training and testing. We can finetune the UAV model for lower resolution airplane data. This is the first step.

Depending on your interest, there are a number of avenues from there. Here is one option depending on time investment and interesting.

• There are many images of the same colony as the photographer flies around it. If we have 30 images of the same colony, can we estimate the total number of unique birds in the colony? Some interesting papers (https://www.mdpi.com/2073-4395/12/2/440, https://arxiv.org/pdf/2012.00946.pdf, https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136690224.pdf)

This would be important for the field of ecological monitoring at large. There are many times we have unstructured photos from a single 'unit', that we desire the total count. There are many ways to go about it, from re-identification, to temporal overlay, to image-matching, etc. A fun place to explore. I have other options as well that we can discuss.

[Mu-Magdy]

Thanks alot @bw4sz,
That was very helpful,

But to make sure I understand correctly, Is the images annotated or I need to make time in the proposal to annotate data? As this 'An annotated dataset for training and testing, contributing to the improvement of the model's performance.' was written in the original project idea file as an expected output

And the part related to bird counting, Is the expected output a model to count the birds in the 'unit' or a comparison to different approaches or papers to choose the best on a later stage? or both is expected

[Mu-Magdy]

Hi @bw4sz,

Hope you are fine,
Can you please share the trained model and any other relevant work to prepare for next steps?

[bw4sz]

The trained model is here: https://www.dropbox.com/scl/fi/e19bll7t1e8aigirj8vtx/species_model.pl?rlkey=culy6hqgma3esc06h49cty6r3&dl=0

It can be opened like so:

from deepforest import main
m = main.deepforest.load_from_checkpoint(model_path)