Fine-grained recognition is where vision models meet classes that differ in sub-millimeter detail, long-tailed label distributions and strong domain shift. I benchmark self-supervised, supervised and vision-language encoders with linear probing, measure sample efficiency, and test transfer from curated images to in-the-wild photographs, using ground beetle species as the test bed.
12 vision and vision-language encoders, linear-probed on about 100K images of 1,769 carabid species from four long-tailed datasets. ViLT reaches 97% genus and 94% species accuracy, half the training data costs little, and lab-to-field transfer drops 37%.
The first version of the benchmark, presented at CV4Animals: where frozen encoders succeed on curated collections, and how sharply they fail once the same species are photographed in the field.
With hundreds to thousands of labels, the regime most science lives in, multimodal LLMs plateau early while vision-only encoders keep improving. On NeWT, 10x more training samples beats 10x more inference compute.