README.md

KinematicNet

Preparing the data

The dataset is preprocessed. The annotation file and the videos can be downloaded from here. Once the data are downloaded, follow some pre-processing of the videos and organize the filesystem like this.

${ROOT}
    |--data
    |   |--images/
    |   |--annotations.npy
    |--dataset
    |...

How to run the code

First, install necessary packages via virtualenv by running:

// create the virtual environment
$ virtualenv venv
// activate the virtual environment
$ source venv/bin/activate
// install required packages
(venv) $ pip3 install -r requirements.txt

Then, run the training code:

// for SO(3) supervision
(venv) $ python3 train_roofing.py --config experiments/roofing/l2/<config>.yaml
// OR for direct supervision
(venv) $ python3 train_roofing_direct.py --config experiments/roofing/l2/<config>.yaml

Multi-GPUs can be run by commenting out the line model = nn.DataParallel(model) in the main() method.

To evaluate the results, run:

(venv) $ python3 evaluate_roofing.py --record logs/<path-to-result-file>.pkl