Ecoset – an ecologically more valid large-scale image dataset

Home » Ecoset – an ecologically more valid large-scale image dataset
Introduction
Tired of all the dogs in ImageNet (ILSVRC)? Then ecoset is here for you. 1.5m images from 565 basic level categories, chosen to be both (i) frequent in linguistic usage, and (ii) rated by human observers as concrete (e.g. ‘table’ is concrete, ‘romance’ is not). Here we collect resources associated with ecoset. This includes the dataset, trained deep neural network models, code to interact with them, and published papers using it.

THE ECOSET DATASET IS AVAILABLE HERE (the password for extraction can be found in the readme).

The ecoset pre-trained networks, and online tools to run them are available here.

Have you used ecoset in your research, and want to make the corresponding paper/resources available here? Let us know, happy to help.

Ecoset in brief
    • 1.5 million images
    • 565 basic level object categories
    • Categories selected to be frequent and concrete
    • Expected error rate of <4%
    • Comes with 40 pre-trained DNN instances (Alexnet, vNet)
    • Easy online tool to extract network activations for new stimuli
    • Main paper available here
    • Dataset available via HuggingFace (password in README_license_ecoset_password.md)
    • Code and networks available here

Ecoset resources
Main Paper introducing ecoset:

DNNs & Code (CodeOcean):

Twitter summary of the ecoset project:

ecoset on HuggingFace: