Implicit Stacked Autoregressive Model for Video Prediction

Project Page | Video | Paper |

Minseok Seo, Hakjin Lee, Doyi Kim, Junghoon Seo,
SI Analytics

---

NEWS

Our typhoon data set and experimental results will be published soon. [2023.06.01]

(This is not included in our report. prediction +10 hours)

Requirements

scikit-image==0.16.1

Moving MNIST

sh data/moving_mnist/download_mmnist.sh
python3 main.py

Pre-trained weight

(We are in the process of reproducing the performance by running the refactored code and updating ckpt.)

Dataset	Epoch	MSE	Weight
Moving MNIST	100	31.04	ckpt
Moving MNIST	200	27.04	ckpt
Moving MNIST	300	24.92	ckpt
Moving MNIST	400	23.56	ckpt
Moving MNIST	500	22.57	ckpt
Moving MNIST	600	21.67	ckpt
Moving MNIST	800	20.50	ckpt
Moving MNIST	1,000	19.53	ckpt
Moving MNIST	1,300	18.57	ckpt
Moving MNIST	1,500	18.03	ckpt
Moving MNIST	2,000	17.20 (paper:16.9)	ckpt

Web Demo

IAM4VP web deomo available at https://ovision.ai/ [it will plan 2023.08.01]

Thank to

This code is heavily based on SimVP.

We thank the authors of that code.

Citation

@article{seo2023implicit,
  title={Implicit Stacked Autoregressive Model for Video Prediction},
  author={Seo, Minseok and Lee, Hakjin and Kim, Doyi and Seo, Junghoon},
  journal={arXiv preprint arXiv:2303.07849},
  year={2023}
}