Data Structure

For the following datasets, please use our package '+gsis' to work with the data structure. You can obtain the package from our ImageAnnotation tool here. Processed datasets are stored in image stores. In general, an image store consists of a vocabulary and a lot of images, each image has two types of information:

  • visual features
  • (optional) textual labels taken from the vocabulary

Information of an image store is specified in one store_info text file. One store consists of several files/folders as follows:

  • words: contains vocabulary
  • images: contains image id
  • rawImageFolder: contains images. Those images can be resized or preprocessed but still in an image format such as Jpeg or Bmp.
  • image_word: contains mapping from imageid to wordid
  • image_feature_folder: mapping from imageid to folder (inside feaRootFolder specified in store_info file) containing features of the image. This is used to reduce the number of files and size of each feature folder.

Due to the huge size of features, we couldn't upload feature folders with dataset. Please use our tool and re-extract features like the instructions provided in the README files of the tool and inside the uploaded zip files of preprocessed datasets

List of the datasets:

  1. Processed UWDB: UWDB is an annotation dataset provided by University of Washington. Please refer to their website for the original dataset. Here, we only provide the processed dataset stored in our data structure [Download]

  2. Flickr dataset: We obtain 5500 images with tags from Flickr using FlickrAPI. For the raw dataset, please download from here. For the processed dataset, click here to download