fengxu-pku / visual-dynamics

Visual Dynamics: Probabilistic Future Frame Synthesis via Cross Convolutional Networks

Visual Dynamics: Probabilistic Future Frame Synthesis via Cross Convolutional Networks

This repository contains pre-trained models and demo code for the project 'Visual Dynamics: Probabilistic Future Frame Synthesis via Cross Convolutional Networks', published at the Conferece of Neural Information Processing Systems (NIPS) 2016.

http://visualdynamics.csail.mit.edu/

Prerequisites

Torch

We use Torch 7 (http://torch.ch) for our implementation.

Imagemagick (optional)

We use the convert toolbox in 'Imagemagick' to generate gif images for visualization. See demo.lua about how to disable it.

Installation

Our current release has been tested on Ubuntu 14.04.

Clone the repository

git clone [e-mail]:tfxue/visual-dynamics.git 

Download pretrained models and sample vectors (85MB)

cd visual-dynamics
./download_models.sh

Run test code

cd src
th demo.lua

There are a few options in demo.lua:

useCuda: Set to false if not using Cuda

gpuId: GPU device ID

demo: Set it to 'all' to run all demos. Set it to 'demo?' to run a specific demo

modeldir, datadir, outputdirRoot: Directories that store models, input files, and output files

createGIF: Generate gif visualizations. This requires Imagemagick. Set it to false if Imagemagick is not installed.

Sample input & output

Demo 1: sample future frames from a single image

Input image	Sample 1	Sample 2	Sample 3

Demo 2: transfer motion from a source pair to a target image

Source motion	Target image 1	Transfered motion	Target image 2	Transfered motion

Demo 3: visualize selected dimensions of the latent representation

Dimension 0752	Dimension 1746	Dimension 2195

Datasets we used

• Exercise dataset: zip, 1.1GB

Reference

@inproceedings{visualdynamics16,   
    author = {Xue, Tianfan and Wu, Jiajun and Bouman, Katherine L and Freeman, William T},   
    title = {Visual Dynamics: Probabilistic Future Frame Synthesis via Cross Convolutional Networks},   
    booktitle = {NIPS},   
    year = {2016}
}

For any questions, please contact Tianfan Xue ([e-mail]) and Jiajun Wu ([e-mail]).