New Brown Corpus

This repository provides data and code to use the New Brown Corpus dataset. To view this dataset online, visit https://lunar.cs.brown.edu/nbc

Dependencies

• numpy
• pandas
• pickle
• tqdm
• argparse

Download

• Clone this repository
• In the project directory, download and extract:
  • https://plunarlabcit.services.brown.edu/nbc/downloads/release.7z (2.6G, required)
  • https://plunarlabcit.services.brown.edu/nbc/downloads/images.7z (311G, optional)
• Set your NBC_ROOT environment variable to this directory. For example on linux, run export NBC_ROOT=/path/to/nbc
• If everything is set up correctly, you should be able to run nbc.py without error

Usage

• Import the NBC class

import sys  
sys.path.append('path/to/nbc_parent_dir')  
from nbc.nbc import NBC

• Create argparse arguments to construct an NBC object.

import argparse
parser = argparse.ArgumentParser()
NBC.add_args(parser)
args = parser.parse_args([
  '--features', 'posX:LeftHand', 'posY:RightHand' posZ:RightHand'
])
dataset = NBC(args)

• Access dataset properties/methods as needed. For example, to access train/test feature matrices:

train_x = np.vstack(list(dataset.features['train'].values()))
test_x = np.vstack(list(dataset.features['test'].values()))

See nbc.py for available properties, methods, and arguments.

Notes

• The dataset will take a while to load for each new subsampling/dynamic_only argument configuration. However, temporary files will be stored in tmp/ to allow faster loading.
• Examples using the image data currently aren't available. However, image data follows the format {session_name}/{step_number}.png, and corresponds one-to-one to the spatial data.