HELI⬤S

Structure of the dataset

HELIOS is a large dataset of individual button presses (>1M) generated by humans in response to noisy images, and in compliance with the specifications of elementary visual tasks. For example, input images may represent different samples of a noise source applied to a target line, which may be absent in some images. Observers may be asked to determine whether the target bar was present or absent. HELIOS provides you with each input image, and each corresponding binary response (absent/present) produced by the observer.

Motivation

HELIOS is intended for modellers interested in detailed characterizations of low-level human behaviour. All datasets are formatted with machine learning applications in mind: a large set of images with associated labels. Downloads also include other information that is project-specific, such as subject ID, contrast, and others. Please refer to the README file for more information.

What this is not

HELIOS is constructed around elementary visual tasks defined on image primitives, such as detection/discrimination of edges and bars. Different images do not contain completely different scenes, like a house in one image and a dog in another. Rather, they represent noisy variations around consistent exemplars, such as an edge embedded in noise sample 1, and the same edge embedded in noise sample 2. Furthermore, output labels do not refer to complex concepts, like object identity or segmentation maps, but rather reflect elementary judgments (absent/present, left/right). For the above reasons, it is fundamentally different from ImageNet, for example, and it targets different aspects of sensory behaviour.

Downloads

Each link below corresponds to a different experimental project, and each project is associated with its own zip file. The zip file contains two directories: images and labels. The images folder contains all images in png format. The labels folder contains various labels in both csv and json format. For csv, each label set is associated with its own csv file. For json, one file contains all labels with corresponding names. The zip folder also contains a README file that provides details about the project, and references to relevant publications. If you use HELIOS, please acknowledge it by citing those publications.

HELIOS01 (~520K trials)

HELIOS02 (~200K trials)

HELIOS03 (~100K trials)

HELIOS04 (~250K trials)

Loaders

Each download includes three data loaders for importing images/labels into:

1) generic python format;

2) generic Matlab format;

3) PyTorch Dataset primitives.

Contact

If you encounter difficulties with the above material, you can contact me at this email address and I will do my best to address the issue.