Cicindela_Platydracus_ID_FMNH

Cicindela + Platydracus Taxonomic Classifier

A multilabel deep-learning model that jointly identifies Cicindela tiger beetles (Cicindelidae) and Platydracus rove beetles (Staphylinidae) from images of pinned museum specimens. The model predicts a hierarchical label set covering genus, species group (Platydracus only), species, and subspecies simultaneously.

Model Details

Built on EVA-02 Large (448 px) from timm with a shared backbone and three classification heads:

• Genus head (softmax, 2 classes): routes inference to the appropriate genus-specific head.
• Cicindela head (sigmoid, 248 classes): species and subspecies labels for Cicindela.
• Platydracus head (sigmoid, 77 classes): species group, species, and subspecies labels for Platydracus.

Training used single-label pretraining followed by multi-label fine-tuning with AsymmetricLoss and square-root inverse-frequency weighted sampling. Mixed precision (fp16) on 2 GPUs.

• Developed by: Bruno A. S. de Medeiros, Negaunee Assistant Curator of Pollinating Insects, Field Museum
• Model type: Multi-head image classification
• License: Apache 2.0
• Base model: eva02_large_patch14_448.mim_m38m_ft_in22k_in1k
• Training code: de-Medeiros-insect-lab/Beetle_taxon_id

Label Vocabulary

327 labels total organized in a four-level hierarchy:

Species groups are defined only for Platydracus (14 morphologically diagnosable groups). See vocab.json for the full label list.

Uses

Direct Use

Identification of pinned Cicindela and Platydracus specimens from drawer-scan images generated by DrawerDissect. See the citation below for full dataset and method details.

Out-of-Scope Use

• Species not represented in the FMNH collection.
• Non-pinned or live specimens.
• Other beetle genera.

Evaluation

Metrics on the held-out test set (specimen-weighted, excluding genus-level labels):

Usage

Load the FastAI learner directly (requires the same fastai environment and the MultiHeadBeetleModel class from the training notebook):

from huggingface_hub import from_pretrained_fastai
learn = from_pretrained_fastai("brunoasm/Cicindela_Platydracus_ID_FMNH")

Inference follows standard fastai conventions. The genus head routes each specimen to the appropriate species-level head automatically.

Training Data

Images of pinned museum specimens generated using DrawerDissect on Field Museum collections. See the citation for details.

Citation

Postema, E. G., Briscoe, L., Harder, C., Hancock, G. R. A., Guarnieri, L. D., Eisel, T., Welch, K., Fischer, N., Johnson, C., de Souza, D., Phillip, D., Baquiran, R., Sepulveda, T., & de Medeiros, B. A. S. (2025). DrawerDissect: Whole-drawer insect imaging, segmentation, and transcription using AI. EcoEvoRxiv. https://doi.org/10.32942/X2QW84

Model Card Author

Bruno de Medeiros, Negaunee Assistant Curator of Pollinating Insects, Field Museum