This repository holds the code used for our recent work on domain randomization. We try to study how we can use this technique in the creation of large domain-specific synthetic datasets. We conclude that this approach can be preferable to fine-tuning state-of-the-art object detection CNNs on small real image datasets, after being pre-trained on huge and available datasets such as COCO.
We provide a Jupyter notebook with a step-by-step walkthrough which includes:
- How to preprocess image dataset and create binary
TF records; - How to train network with several configurations (e.g. starting checkpoint, training set, learning rate);
- How to get mAP (Mean Average Precision) curves on validation set over time;
- How to export inference graphs and infer detections;
- How to obtain AP metrics and precision-recall curves on test set.
The tests we performed are summarized in tests.md.
We have uploaded a preprint version of our work to arXiv.org which can be cited as follows:
@article{borregoatabak2018,
title={The impact of domain randomization on object detection: a case study on parametric shapes and synthetic textures},
author={Dehban, Atabak and Borrego, Jo{\~a}o and Figueiredo, Rui and Moreno, Plinio and Bernardino, Alexandre and Santos-Victor, Jos{\'e}},
booktitle="IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)",
year={2018}
}
We provided the used real image dataset in our laboratory's webpage. The used synthetic datasets were generated using an open-source Domain Randomization plugin for Gazebo simulator. A detailed walkthrough on how to generate such datasets is provided here.
Our pipeline is based on Justin Francis' TensorFlow Object Detection API repository.
Our pipeline requires the following software.