N3: Newtonian Image Understanding: Unfolding the Dynamics of Objects in Statis Images
This is the source code for Newtonian Neural Networks N3, which predicts the dynamics of objects in scenes.
Citation
If you find N3 useful in your research, please consider citing:
@inproceedings{mottaghiCVPR16N3,
Author = {Roozbeh Mottaghi and Hessam Bagherinezhad and Mohammad Rastegari and Ali Farhadi},
Title = {Newtonian Image Understanding: Unfolding the Dynamics of Objects in Static Images},
Booktitle = {CVPR},
Year = {2016}
}
Requirements
This code is written in Lua, based on Torch. If you are on Ubuntu 14.04+, you can follow this instruction to install torch.
You need the VIND dataset. Extract it in the current directory, and rename it to VIND
. Or you can put it somewhere else and change the config.DataRootPath
in setting_options.lua
.
Training
To run the training:
th main.lua train
This trains the model on training data, and once in every 10 iterations, evalutates on one val_images
batch. If you want to validate on val_videos
go to setting_options.lua
and change the line valmeta = imvalmeta
to valmeta = vidvalmeta
.
Test
You need to get the weights. Extract the weights in the current directory and rename it weights
. To run the test:
th main.lua test
License
This code is released under MIT License.