This package provides annotations and a simple toolkit for the Novel Visual Concept (NVC) dataset.
Please run setup.sh to download images and pre-calculated VggNet layer 15 image features.
For more details, please run NVC_dataset_demo.ipynb. You are also welcome to look into the python class in ./python_lib/class_nvc_dataset.py
If you get a "bad request" error message when opening NVC_dataset_demo.ipynb, you probably need to update your ipython. You can also run NVC_dataset_demo.py instead.
If you find this dataset and annotations useful in your research, please consider citing:
@inproceedings{mao2015learning,
title={Learning like a Child: Fast Novel Visual Concept Learning from Sentence Descriptions of Images},
author={Mao, Junhua and Xu, Wei and Yang, Yi and Wang, Jiang and Huang, Zhiheng and Yuille, Alan},
booktitle={ICCV},
year={2015}
}
This annotations contains two JSON files (in ./annotations): one for training and validation set, one for testing set. The JSON file is organized as follows:
The root is a key-value dictionary:
- 'version': version of the dataset
- 'concepts': a list of the novel visual concepts
- 'images': a list, each element of the list is a dictionary:
- 'concept': novel concepts for the image
- 'image_id': unique id for the image
- 'image_name': file name for the image
- 'train_val_test_split': 'train' or 'val' or 'test'
- 'sentences': a list, each element of the list is a dictionary:
- 'raw': a string which is the raw annotated sentence
- 'tokens': tokenized sentence, NOT includes the period at the end
- 'sentence_id': unique id for the sentence
- 'image_id': unique image id that the sentence belongs
The annotations in this dataset belong to University of California, Los Angeles and Baidu Research, and are licensed under a Creative Commons Attribution 4.0 License.