Novel Visual Concept (NVC) dataset
收藏资源简介:
该数据集提供了新颖视觉概念的标注和工具包,用于研究从图像句子描述中快速学习新颖视觉概念。数据集包含训练、验证和测试集的JSON文件,每个文件详细记录了图像和相关句子的信息。
This dataset provides annotations and toolkits for novel visual concepts, designed for researching the rapid learning of novel visual concepts from image-sentence descriptions. The dataset includes JSON files for training, validation, and test sets, each meticulously documenting the information of images and their associated sentences.
Novel Visual Concept (NVC) Dataset Summary
Overview
The Novel Visual Concept (NVC) dataset provides annotations and a simple toolkit for research purposes. It includes images and pre-calculated VggNet layer 15 image features, which can be downloaded by running setup.sh.
Dataset Structure
- Annotations: Two JSON files located in the ./annotations directory:
- One for the training and validation set.
- One for the testing set.
JSON File Structure
- Root: A key-value dictionary containing:
- version: Dataset version.
- concepts: List of novel visual concepts.
- images: List of dictionaries, each containing:
- concept: Novel concepts for the image.
- image_id: Unique image identifier.
- image_name: File name of the image.
- train_val_test_split: train, val, or test.
- sentences: List of dictionaries, each containing:
- raw: Raw annotated sentence.
- tokens: Tokenized sentence without the period.
- sentence_id: Unique sentence identifier.
- image_id: Unique image identifier the sentence belongs to.
License
The annotations in the dataset are licensed under a Creative Commons Attribution 4.0 License. The dataset is a collaboration between the University of California, Los Angeles, and Baidu Research.




