Hi everyone,

I wanted to share a project I've been working on and get some advice from this community. I developed PantheraID, an individual jaguar identification system based on an ensemble of convolutional neural networks, which reaches 96.6% precision on Costa Rican data and over 90% on an independent Brazilian dataset. It's deployed in production, and an article about the project is currently being prepared for publication in the journal Methods in Ecology and Evolution.

The source data comes from camera trap videos collected by two NGOs in Costa Rica, gathered over 14 years of continuous monitoring.

The goal is automated identification of individual jaguars. Today this process is done by a handful of field experts, who have to manually classify each record alongside all their other responsibilities, which takes a lot of time. Knowing which jaguar is which matters because, by telling individuals apart, you can count how many there are, and this serves as an indicator of ecosystem health. The system could also be applied to other big cats with unique coat patterns.

Right now I'm looking for some advice. I'm currently manually cleaning my original dataset, keeping only a few representative frames per video (instead of every extracted frame), to avoid having near-identical frames from the same event split across training and validation. My theory is that this may have contributed to a memorization/overfitting issue I detected earlier, but I haven't fully confirmed that yet, it's a hypothesis I'm still testing.

The side effect is that my dataset is now much smaller, which is a classic challenge with jaguars and big cats in general: they're hard to photograph well, and there are few events per individual.

If anyone has dealt with similar dilemmas (balancing data cleaning against dataset size, individual re-identification with few examples, data leakage in camera trap datasets), I'd love to hear how you approached it. And if this is something you're interested in and would like to think through it with me, you're very welcome to reach out.

Feel free to email me at [email protected]/pantheraid Thanks!