S M Rayeed
Home / Fine-grained taxonomic identification
ICCV 2025 WorkshopFirst author

Fine-Grained Beetle Taxonomy with Vision Models: A Benchmark on Long-Tailed and Domain-Adaptive Classification

S M Rayeed, Alyson East, Samuel Stevens, Sydne Record, Charles V. Stewart

ICCV 2025 Workshop on Computer Vision for Ecology (CV4E)

PaperSupplementarXiv (CVPRW version)Cite
Four long-tailed datasetsBeetlePaloozaLab, ethanol vialsBeetlePUUMLab, pinned, HawaiiNHM-CarabidsMuseum, LondonI1MCField photographs100,885 images230+ genera, 1,769 species12 frozen encodersVision, self-supervisedDINOv2, ViT-MAE, SwAV, MoCo v3Vision, supervisedSwin V2, BEiT, LeViT, ConvNeXtVision and languageCLIP, BioCLIP, SigLIP, ViLTLinear probegenus and species headsBest: ViLT + MLP97.15%genus93.97%speciesZero-shot BioCLIP3.21% speciesHalf the training data,almost the same accuracyLab to field transfer:37% performance drop
Four long-tailed carabid datasets are embedded by 12 frozen encoders and classified with linear probes at genus and species level. ViLT with an MLP head is the strongest model.

A probe on a frozen ViLT names ground beetle species with 94% accuracy, needs about half the labels, and still loses 37% when the photos come from the field.

Ground beetles (Carabidae) are sensitive indicators of habitat change, and programs such as NEON trap them across the continent every year. Identifying them is the bottleneck: species differ in sub-millimeter details of the elytra and pronotum, and only trained taxonomists can do it reliably.

This benchmark asks which modern vision encoders can take over that task. We evaluate 12 vision-only and vision-language models on four datasets that span pinned museum specimens, ethanol-preserved vials and field photographs, and test them in the two settings that decide whether automation is practical: limited labels and a change of imaging domain.

Benchmark at a glance

Images100,885
Taxa230+ genera, 1,769 species, strongly long-tailed
DatasetsBeetlePalooza (NEON vials), BeetlePUUM (Hawaii, pinned), NHM-Carabids (Natural History Museum, London), I1MC (Insect-1M field images of Carabidae)
EncodersSelf-supervised: DINOv2, ViT-MAE, SwAV, MoCo v3. Supervised: Swin V2, BEiT, LeViT, ConvNeXt. Vision and language: CLIP, BioCLIP, SigLIP, ViLT
ProtocolFrozen embeddings, linear and MLP probes at genus and species level; BioCLIP zero-shot as the baseline
Accuracy on merged data (%)025507510023.297.278.3Genus3.294.065.7SpeciesBioCLIP zero-shotViLT + MLP probeViLT macro accuracy
Zero-shot BioCLIP reaches 77% at family level but falls to 3% at species level. A trained probe on ViLT reaches 94%. Macro accuracy averages over classes, so rare species count as much as common ones.

Findings

  1. Vision-language pretraining helps under probing. ViLT with an MLP head reaches 97.15% genus and 93.97% species accuracy on the merged data, with 78.30% and 65.67% macro accuracy across classes.
  2. Zero-shot is not enough for taxonomy. BioCLIP, built for the tree of life, gets 77.37% at family but 23.17% at genus and 3.21% at species without training.
  3. Labels can be halved. Subsampling the training set to 50% leaves accuracy largely unchanged, which lowers the annotation cost for long-tailed collections.
  4. The field is the hard part. Models transfer well between curated lab collections, but performance falls by 37% on in-situ images, where pose, occlusion and background vary.
Curated collectionsBeetlePaloozaBeetlePUUMNHM-CarabidsTransfer holds between lab collectionsIn the fieldI1MC: varied pose, occlusion, cluttertrain37%performance drop fromlab-trained modelsViLT + MLP: 97% genus, 94% species accuracy on merged data.50% of the training set gives nearly the same accuracy.
Transfer between curated collections holds; transfer from lab images to field photographs does not.

The same protocol is now being extended beyond beetles to butterflies and moths, to test which conclusions hold across taxa.

Related work

  1. CVPR 2025 WorkshopFirst author

    BeetleVerse: A study on taxonomic classification of ground beetles

    S M Rayeed, Alyson East, Samuel Stevens, Sydne Record, Charles V. Stewart
    CVPR 2025 Workshop on CV4Animals
    PaperarXivPoster
  2. NeurIPS 2025 Workshop

    Mind the (Data) Gap: Evaluating Vision Systems in Small Data Applications

    Samuel Stevens, Jenna Kline, S M Rayeed
    NeurIPS 2025 Workshop on Imageomics
    PaperarXivCode
  3. Scientific Data 2026First author

    A continental-scale dataset of ground beetles with high-resolution images and validated morphological trait measurements

    S M Rayeed, Mridul Khurana, Alyson East, Isadora E. Fluck, Elizabeth G. Campolongo, Samuel Stevens, Iuliia Zarubiieva, Scott C. Lowe, Michael W. Denslow, Evan D. Donoso, Jiaman Wu, Michelle Ramirez, Benjamin Baiser, Charles V. Stewart, Paula Mabee, Tanya Berger-Wolf, Anuj Karpatne, Hilmar Lapp, Robert P. Guralnick, Graham W. Taylor, Sydne Record
    Scientific Data (Nature Portfolio), Data Descriptor
    Project pagePaperDOIData: Hawaii-beetlesData: 2018-NEON-beetlesCode

Cite

@inproceedings{rayeed2025finegrained,
  title     = {Fine-Grained Beetle Taxonomy with Vision Models: A Benchmark on Long-Tailed and Domain-Adaptive Classification},
  author    = {Rayeed, S M and East, Alyson and Stevens, Samuel and Record, Sydne and Stewart, Charles V.},
  booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
  year      = {2025}
}