July 13, 2026 Dataset Open Access

Metric-Guided Synthetic Image Data Rendering for Deep Learning compatible with Agentic AI

Radoynova, Martina; Pantze, Samuel; De, Trina; Günther, Ulrik; Yakimovich, Artur

This repository contains the datasets, 3D rendering scene files, and trained model weights accompanying the manuscript "Metric-Guided Synthetic Image Data Rendering for Deep Learning compatible with Agentic AI." The deposited files provide the complete procedural pipeline and resulting artefacts used to evaluate zero-shot instance segmentation of virological plaque assays (based on the VACVPlaque dataset).

Repository Contents

• datasets/ Contains the synthetic image datasets generated at three varying levels of perceived realism (High, Medium, Low) and in two sizes (Small [S]: 100 images; Large [L]: 1,000 images). It also includes the Mix dataset, which consists of 90 high-realism synthetic images and 10 real images.

• blender/ Includes the procedural 3D scene files used to generate the synthetic datasets via Blender. These files can be used manually or coupled with the SynthClaw agentic skill for automated rendering.

  • Scene_High.blend

  • Scene_Medium.blend

  • Scene_Low.blend

• model_weights/ Contains the trained single-shot architecture model weights and threshold configurations used to evaluate the zero-shot performance of each generated dataset.

  • Evaluated Conditions: High_L, High_S, Medium_L, Medium_S, Low_L, Low_S, and Mix.

  • Files per Condition: Each subdirectory includes the best and last model weights (weights_best.h5, weights_last.h5) alongside the corresponding threshold optimization parameters (thresholds1.json, thresholds2.json) for large and small object segmentation.