Synthetic Pallet Datasets for Fine-Tuning Faster R-CNN (Realistic / Half-Realistic / Random)

Synthetic Pallet Datasets for Fine-Tuning Faster R-CNN This archive provides three synthetic datasets used to study the impact of realism and domain randomization on pallet detection when fine-tuning a pre-trained Faster R-CNN. Each dataset corresponds to a single generation recipe – realistic, half-realistic, or random – and is packaged as a separate ZIP file. ZIP files: realistic.zip, halfrealistic.zip, random.zipInside each ZIP: four splits test/ train_10_percent/ train_50_percent/ train_100_percent/ Each split contains: annotations/ – ground-truth bounding box files (one file per image). images/ – RGB images used for training/evaluation. (train_ only)* Best model weights per seed as separate .pth files. (train_ only)* Training logs and the corresponding in-distribution (ID) and out-of-distribution (OOD) test performance after fine-tuning on that split. (train_ only)* Per-seed performance visualizations. Intended Use. Reproduce our experiments, compare training strategies (mixed vs. bridged transfer), or benchmark alternative detectors in low-data settings. Source / Creation Process. Images were generated with NVIDIA Omniverse Replicator using three variant recipes (Realistic / Half-Realistic / Random). Ground-truth boxes were produced during rendering; evaluation splits mirror those in the paper.File Organization (Example)random.zip ├─ test/ │ ├─ annotations/ │ └─ images/ ├─ train_10_percent/ │ ├─ annotations/ │ ├─ images/ │ ├─ *_ood_test_log_seed_*.txt │ ├─ *_test_log_seed_*.txt │ ├─ *_training_log_seed_*.txt │ ├─ *.png per seed │ └─ best *.pth per seed ├─ train_50_percent/ (same structure) └─ train_100_percent/ (same structure) Related Resources Project Page: https://muammerbay.github.io/omniverse-replicator-sim2real-analysis/ GitHub: https://github.com/MuammerBay/omniverse-replicator-sim2real-analysisarXiv: https://arxiv.org/abs/2510.12208 How to Cite @misc{bay2025impactsyntheticdataobject, title={The Impact of Synthetic Data on Object Detection Model Performance: A Comparative Analysis with Real-World Data}, author={Muammer Bay and Timo von Marcard and Dren Fazlija}, year={2025}, eprint={2510.12208}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2510.12208}, }