Who you work with#
Atria Foundation, the charitable arm of Atria Group, described in Brief 12. One of its four pillars is Bio-Diversity, which it describes as promoting harmony across species. Its visible work there so far has been plantation drives.
This problem sits further from that existing work than the Foundation’s other briefs, and it has one hard dependency: identifying individual animals needs a photographic catalogue that somebody has already built. Where that comes from is the first thing to settle.
The situation#
Wildlife management asks questions about individuals, not just species. Is one elephant responsible for most of the crop raiding in a valley, or fifty? Is the animal we translocated the one that came back? How many distinct individuals use this corridor?
Right now those questions are answered by researchers who know the animals personally, which does not scale and disappears when they move on.
What is going wrong#
Without individual identification, conflict data describes a place rather than an animal. That changes what you can do about it. A single repeat individual is a management problem with a specific answer. Fifty animals passing through is a landscape problem with a different one. The data cannot currently tell you which you have.
Why it is not fixed yet#
Some species make this easy. Tigers have stripe patterns that work almost like barcodes. Elephants do not. Identification relies on a combination of features that are harder to read from a photograph: ear fold shape, notches and tears in the ear edge, depigmentation patches, tail and tail-brush condition, body scars, and tusks where present.
There is also a harder statistical problem underneath, which most projects skip. A system that matches every photograph to the closest animal in its catalogue will confidently mislabel an animal it has never seen. Recognising that a new individual is new is the part that matters and the part that is difficult.
Skills and Expertise#
- Branch — CS and Electronics. Strong computer vision content.
- Technical skills — Computer vision and re-identification, metric learning and embeddings, open-set recognition, working with small and imbalanced datasets, human-in-the-loop system design.
- Other skills — Reading and building on published research, scientific writing, reporting error honestly, handling sensitive data.