/DeepOIS

[T-CSVT 2021]: DeepOIS: Gyroscope-Guided Deep Optical Image Stabilizer Compensation

Primary LanguagePythonMIT LicenseMIT

DeepOIS: Gyroscope-guided deep optical image stabilizer compensation

How to generate GT labels

We introduce Essential-Mixtures Model to calculate GT labels, which is effective. However, it requires non-trivial computation resources and is not stable. So we here provide an alternative method called Homography Mixtures.

Model and Data

The pretrain-model, benchmark, trainset could be found at google-drive

  • Download and Put GF4-Val.txt GF4-Val to the root path
  • Download and Unzippretrain-model to the root path

Here is the colab demo for illustrating the GF4 benchmark

Inference

sudo python3 test_demo.py --data_dir TestDatasetItem.npy --model_dir pretrain-model --restore_file best

Training

coming soon

Citation

@article{liu2021deepois,
  title={DeepOIS: Gyroscope-guided deep optical image stabilizer compensation},
  author={Liu, Shuaicheng and Li, Haipeng and Wang, Zhengning and Wang, Jue and Zhu, Shuyuan and Zeng, Bing},
  journal={IEEE Transactions on Circuits and Systems for Video Technology},
  volume={32},
  number={5},
  pages={2856--2867},
  year={2021},
  publisher={IEEE}
}