This repository is the official implementation of A Scaling Law for Synthetic-to-Real Transfer: A Measure of Pre-Training.
This repository contains the code to generate synthetic images, the results for reproducing the plots, and pre-trained models.
To generate synthetic images/annotations we used in the experiments, see rendering directory.
To check the accuracy scores of various pre-training/fine-tuning task pairs, see results directory.
We provide pre-trained backbone networks used in the paper. Here, task
means the pre-training task, data
means the pre-training dataset (see rendering for more details), and # of examples
means the size of the dataset. All the pre-trained models are compatible with resnetxx
of torchvision.
task | data | # of examples | backbone | download |
---|---|---|---|---|
object detection | bop | 64k | ResNet50 | N/A |
multiclass classification | bop | 64k | ResNet50 | N/A |
surface normal estimation | bop | 64k | ResNet50 | N/A |
semantic segmentation | bop | 64k | ResNet50 | N/A |
Due to the filesize, we cannot put the download links here. If you are interested in, please let me know.
If you cite our work, please use the following bibtex entry.
@article{mikami2021a,
title={A Scaling Law for Synthetic-to-Real Transfer: A Measure of Pre-Training},
author={Hiroaki Mikami and Kenji Fukumizu and Shogo Murai and Shuji Suzuki and Yuta Kikuchi and Taiji Suzuki and Shin-ichi Maeda and Kohei Hayashi},
year={2021},
journal={arXiv preprint arXiv:2108.11018}
}