GeoComb: supporting data, model weights, and Windows applications

Abstract

GeoComb is a desktop application for detecting honeycomb cells and classifying their contents in frame photographs. This repository provides supporting data, trained model weights, analysis scripts, and Windows applications for the GeoComb study. The data archive includes 52 in-house frame images and annotations, divided into 39 training, 7 validation, and 6 test images; ROI annotation records and dataset inventories; and reference annotations, saved predictions, and analysis data for the 29-image benchmark comprising 13 DeepBee, 10 FAIRHive, and 6 in-house images. Supporting analyses cover geometric detection stages, frame-level cell counts and content composition, and comparison of the hierarchical classifier with a single eight-class classifier. Scripts and figure source data are included for reproducing the reported analyses. Model files include the ROI segmentation model, three hierarchical classification models, and the single eight-class baseline model. Windows CPU and NVIDIA CUDA applications are provided as separate archives. Data package version 1.0.1 updates dataset documentation and adds supporting analysis materials. The four deployment models and reported evaluation values are unchanged. The Windows applications accompany GeoComb software version 1.0.0. DeepBee and FAIRHive source images are available from their original providers. DeepBee-derived training crops and GeoComb-corrected DeepBee training annotations are not redistributed. Source code: https://github.com/zlddh95-eng/GeoComb Licensing: Original images, annotations, analysis results, and documentation created by the GeoComb authors are licensed under CC BY 4.0. GeoComb source code is licensed under AGPL-3.0-only. Third-party data, model components, and bundled dependencies retain their respective licenses and notices.