What Is MiewID?
|
MiewID is Wild Me Lab’s state-of-the-art AI model for visual mark-recapture: identifying individual animals from photographs across marine and terrestrial species. It is the primary matching algorithm in Wildbook.
MiewID learns what makes images of the same animal similar, and images of different animals different. It turns each photo into an embedding, a compact numerical fingerprint of that animal’s unique markings. New photos are matched by comparing their embeddings with those already in a catalog, so new individuals can be added without retraining the model. MiewID can also show the features behind a match, so researchers can check its suggestions. MiewID improves through community collaboration: it is trained on data pooled from public and private mark-recapture catalogs across the Wildbook community. |
Heatmaps highlight the features MiewID uses to match two photos of the same lion
|
MiewID by the Numbers
~90species and 110 feature classes in the v4 training data
78%average top-1 accuracy on withheld validation data (v4)
89%average top-10 accuracy (v4)
1.28M+MiewID matching jobs run in Wildbook
MiewID v4
MiewID v4 is a big leap in training data, species coverage, and performance. Compared with v3, it was trained on 40% more data, covering about 90 species and 110 feature classes such as heads, flanks, fins, flukes, and faces. New species and catalogs in the v4 training data include African and Asian elephants, chital and sika deer, tiger sharks, sand tiger sharks, southern right whales, bowhead whales, fire salamanders, yellow-bellied toads, leafy and weedy seadragons, and northern and southern giraffes.
Averaged across all species, v4 reached 78% top-1, 87% top-5, and 89% top-10 accuracy on withheld validation data, improvements of 2, 3, and 4 points over v3. Because it has learned from so many body shapes and patterns, MiewID can also match some species it has never seen. That “zero-shot” ability helped extend Wildbook support to 25 species of small, understudied wild cats in Whiskerbook. The current release, v4.1, covers 93 species.
Averaged across all species, v4 reached 78% top-1, 87% top-5, and 89% top-10 accuracy on withheld validation data, improvements of 2, 3, and 4 points over v3. Because it has learned from so many body shapes and patterns, MiewID can also match some species it has never seen. That “zero-shot” ability helped extend Wildbook support to 25 species of small, understudied wild cats in Whiskerbook. The current release, v4.1, covers 93 species.
Spotted hyenas
|
Humpback whale flukes
|
Gray whales, from the air
|
Belugas
|
MiewID In-depth
|
MiewID In-depth is a hybrid algorithm that looks at both coarse and fine features. First, MiewID v4.1 makes a fast pass to predict an animal’s identity. Then the LightGlue feature-matching algorithm re-ranks MiewID’s top 50 candidates by fine-grained texture and position, similar to how HotSpotter works. The balance between the two can be tuned for each species.
In testing, 69 of the 93 species in the MiewID v4.1 training data improved. Some of the biggest top-1 gains were juvenile fire salamanders (+31%, to 94%), trout (+19%), Iberian lynx (+16%), chital deer (+15%), ocelots (+13%), and spotted hyenas (+11%). MiewID v4.1 remains the primary algorithm; In-depth is an optional “look again” second pass. |
Top-1 accuracy by species: MiewID In-depth (orange) vs. MiewID v4.1 (blue)
|
Matches in Seconds, With Explanations
|
Wildbook 10.11 rebuilt how MiewID matching works by adding a fast vector database directly into Wildbook. Matching against large catalogs now takes seconds. A single Wildbook that used to handle 3,000–4,000 matching jobs a day can now handle hundreds of thousands.
AI matching tools are often poor at explaining their predictions. Wildbook 10.11 adds PairX, a visualizer developed with the Beery Lab at MIT. It uses matching colors and positions on the two images to show why the AI thinks they are the same animal, so you can check a suggested match quickly. Read the PairX paper. |
PairX showing why two elephant photos match
|
Models, Code, and Research
MiewID models are only licensed for use in Conservation X Labs-hosted applications.
|
Research
• Otarashvili, L., Subramanian, T., Holmberg, J., Levenson, J. J., & Stewart, C. V. (2024). Multispecies animal re-ID using a large community-curated dataset. arXiv. • More publications using Wild Me technologies Documentation • How MiewID matching works in Wildbook |
