Synthetic Training Data from a Virtual Light Trap for Greenhouse Insect Detection: Blender application, reference images, 3D models, training datasets, real captures, and trained models

This dataset accompanies the paper "Synthetic Training Data from a Virtual Light TrapEnables Insect Detection for Greenhouse Pest Management Without Real Images." It providesthe full material needed to reproduce the study's synthetic-data workflow and results. A procedural Blender "virtual light trap" (a digital twin of a physical greenhouse lighttrap) converts a few 2D reference photographs of an insect into large, automatically andexactly labelled synthetic YOLO training sets, via AI-based image-to-3D reconstruction(SAM 3D), multi-pass compositing for error-free ground truth, and targeted domainrandomization of the trap's LED illumination. A lightweight YOLOv26-Nano detector trained onthis synthetic data alone detects real greenhouse insects and runs in real time on edgehardware (NVIDIA Jetson). The deposit covers four greenhouse-relevant study species: the pest whitefly Trialeurodesvaporariorum and the beneficial insects Dacnusa sibirica, Encarsia tricolor, andMacrolophus pygmaeus.