WildID › When a name is wrong

When a species name is wrong

WildID files the answer that BioCLIP gives. It adds no judgement of its own and never changes an answer. A wrong name therefore points to the model or to a setting, and this page shows how to tell which.

Who decides

WildID does four things with a photo. It reads it, it crops to the animal if the crop is on, it asks BioCLIP which names fit, and it files the top answer as keywords and a collection. The choice of the name is the model’s alone.

So a wrong name is rarely a fault in WildID, and WildID cannot repair it. What you can change are the conditions under which the model is asked.

What to check, in this order

  1. Was the animal found? With the crop on, the model sees the animal. With it off, or when nothing is detected, it sees the whole frame, and a small animal in a wide landscape is easily named as a plant or a bird. If several animals are in the frame, only the most confident detection is used.
  2. Was a place list used, and the right one? Without a list the model chooses among 867,455 names from the whole world, look-alikes from other continents included. With the wrong list the right species may not be a candidate at all. Zoo animals and introduced species are usually not on the list of the country they were photographed in.
  3. Was it the right model? BioCLIP 2.5 does not know reptiles and answers with a confident wrong name. Use BioCLIP 2 for reptiles.
  4. Do the taxa fit? A place list covers mammals, reptiles, amphibians and birds unless you change the taxa. An insect or a fish identified with a place list active is forced into one of those groups, and is then usually flagged for review.
  5. Is it a hard case? Closely related species, young animals, unusual angles, motion blur and poor light are difficult for the model, as they are for people.

What the review flag means

WildID flags a photo for review in two cases: the best score is below the threshold, 0.5 unless you change it, or the answer the model would give without the place list is not on the list. Flagged photos get the keyword WildID › review and appear in the _review collection.

The flag is a hint, not a verdict. Flagged photos are often right, and unflagged photos can be wrong.

How often it was right in our tests

These results come from one photographer’s own pictures, taken on a few trips, with the crop on. Lemurs are a hard group with many similar species. Your results will differ.

PhotosPlace listBioCLIP 2BioCLIP 2.5
401 lemur RAW photos, four groupsMadagascar77.1 %80.8 %
The same 401none69.8 %79.8 %
385 lemur JPEGs, twelve speciesMadagascar70.1 %72.5 %
The same 385, genus correctMadagascar90.4 %89.9 %
65 fossa, mongoose and fruit bat photosMadagascar63 of 6565 of 65
22 Kenyan mammal photosKenya21 of 2219 of 22
63 Madagascar reptile photosMadagascar63 of 630 of 63
15 British seabird and eagle photosnone14 of 1514 of 15
A photo counts as correct at the level its folder was labelled with: the species for most, the genus or the family for a few.

Two things stand out. The place list lifted BioCLIP 2 by 4 to 7 percentage points on the lemurs. And the genus was right for nine lemur photos out of ten, so most wrong species names were near misses.

What improves the answers

  • Leave the crop on. In one of our Kenya photos a distant cheetah was named as a francolin, a bird, with a score of about 0.2 on the full frame. With the crop the answer was cheetah at about 0.9.
  • Use a place list for the country or region of the trip. How place lists work
  • Choose the model to suit the subject. BioCLIP 2 or 2.5
  • Switch on the second opinion for mammals. A disagreement between the two models is a good reason to look closer.
  • Go through the review collection before you rely on the names.

What to report, and where

A wrong name is not a bug in WildID. The model’s authors publish new versions from time to time, and WildID can adopt them.

Faults in WildID are worth reporting: a crash, a photo that fails to load, keywords or collections written wrongly, an engine that does not start. Please report those on GitHub.

Previous: GPU, VRAM and speedNext: Place lists and GPS