Simple Shapes Dataset

<p>This dataset is used in the paper <a href="https://arxiv.org/abs/2306.15711"><em>Semi-supervised Multimodal Representation Learning through a Global Workspace, </em>Devillers et al., 2023 (under review)</a>.</p> <p>To use this dataset, use the code provided here: <a href="https://github.com/bdvllrs/bimGW">https://github.com/bdvllrs/bimGW</a>.</p> <p>It consists of 32x32 pixel images of shapes with multiple attributes (size, location, rotation, color). Each image is also paired with its ground truth information (attributes), and a natural language description (English) of the image.</p> <p>The dataset is composed of:</p> <ul> <li>a train set of 500,000 samples,</li> <li>a val and a test set of 1000 samples each.</li> </ul> <p>It also contains already processed 12-dimensional visual features (from a VAE), and presaved BERT features of the text descriptions.</p>