ManyNames is a large scale resource of human object naming data. ManyNames provides on average 31 name annotations for 25K objects in real world images. The images for ManyNames were selected from VisualGenome. In each image, one target object is marked with a red bounding box. Annotators were asked to provide the first name that came to mind for the object tightly bound by this bounding box.
More information and download
The dataset can be downloaded from the GitHub repository. The dataset is provided in .tsv (tab-separated values) and .json format. Python and R scripts to facilitate access to the dataset as well as detailed documentation of the dataset is provided in the repository.
Links to the individual images for the ManyNames version (with bounding box) and the original VisualGenome version (without bounding box) are provided in the dataset. The ManyNames images can also be directly accessed here.
All download options are listed in the download section.
Silberer, C., S. Zarrieß, M. Westera, G. Boleda. 2020. Humans Meet Models on Object Naming: A New Dataset and Analysis. COLING 2020.
Silberer, C., S. Zarrieß, G. Boleda. 2020. Object Naming in Language and Vision: A Survey and a New Dataset. LREC 2020.