Theia: An end-to-end edge-AI mobile framework for high-throughput field phenotyping and on-device geometric morphometrics
Citable archive of Theia accompanying the paper accepted in Methods in Ecology and Evolution (MEE-26-04-354): "Theia: An end-to-end edge-AI mobile framework for high-throughput field phenotyping and on-device geometric morphometrics". This version is the complete research compendium requested by the journal's data archiving policy. It contains, in one record, the source code, the trained models and the annotated dataset that the paper reports on. The GitHub repository (cristobalbragagnolo/Theia) remains the development mirror. Contents cristobalbragagnolo/Theia-v1.2.1.zip: source tree at tag v1.2.1 (Flutter/Dart mobile application, Python deep-learning training pipeline, validation scripts, README with the reproducibility guide). Includes the 268 held-out validation specimens (data/validation_datates/, five populations) with their manual tpsDig ground truth and the population-level morphometric CSV. yolo_dataset.zip: annotated detection and pose dataset, 1,190 Erysimum flower images (one per specimen) with bounding boxes and K = 32 landmarks in YOLO format, split 70/15/15 by specimen (833 train / 178 val / 179 held-out test, random seed 42). detector_nano_fp32.tflite: Stage-1 detector (YOLOv8-Nano), on-device TensorFlow Lite export shipped in the app. pose_medium_lowaug_fp32.tflite: Stage-2 pose model (YOLOv8-Medium-Pose, canonical low-augmentation variant), on-device TensorFlow Lite export shipped in the app. det_baseline_best.pt, pose_lowaug_best.pt: PyTorch weights of the two models above. pose_baseline_best.pt: standard-augmentation ablation variant of the pose model reported in the paper. CHECKSUMS.sha256: SHA-256 digests of the six model and dataset files. Reported performance (held-out test split): detector mAP50-95 = 0.964; pose mAP50-95 = 0.522. Changes since v1.2.1: no code changes. The dataset and model weights, previously distributed only as GitHub Release assets (v1.2.0), are now archived in this record so that the whole compendium carries a persistent DOI. Recommended citation: Bragagnolo, C., Ferreira-Rodríguez, A., Abdelaziz, M., & Muñoz-Pajares, A. J. (2026). Theia: An end-to-end edge-AI mobile framework for high-throughput field phenotyping and on-device geometric morphometrics [Software and data set]. Zenodo. https://doi.org/10.5281/zenodo.19279346