MIM is a neural network for video prediction and spatiotemporal modeling. It is based on the paper Memory In Memory: A Predictive Neural Network for Learning Higher-Order Non-Stationarity from Spatiotemporal Dynamics to be presented at CVPR 2019.

Natural spatiotemporal processes can be highly non-stationary in many ways, e.g. the low-level non-stationarity such as spatial correlations or temporal dependencies of local pixel values; and the high-level non-stationarity such as the accumulation, deformation or dissipation of radar echoes in precipitation forecasting.

We try to stationalize and approximate the non-stationary processes by modeling the differential signals with the MIM recurrent blocks. By stacking multiple MIM blocks, we could potentially handle higher-order non-stationarity. Our model achieves the state-of-the-art results on three spatiotemporal prediction tasks across both synthetic and real-world data.

All pre-trained MIM models have been uploaded to DROPBOX and BAIDU YUN (password: srhv).

It also includes our pre-processed training/testing data for Moving MNIST, Color-Changing Moving MNIST, and TaxiBJ.

For Human3.6M, you may download it using data/human36m.sh.

@article{wang2018memory,
  title={Memory In Memory: A Predictive Neural Network for Learning Higher-Order Non-Stationarity from Spatiotemporal Dynamics},
  author={Wang, Yunbo and Zhang, Jianjin and Zhu, Hongyu and Long, Mingsheng and Wang, Jianmin and Yu, Philip S},
  journal={arXiv preprint arXiv:1811.07490},
  year={2019}
}