In object recognition research, many commonly used datasets (e.g., ImageNet and similar) contain relatively sparse distributions of object instances and views, e.g., one might see a thousand different pictures of a thousand different giraffes, mostly taken from a few conventionally photographed angles. These distributional properties constrain the types of computational experiments that are able to be conducted with such datasets, and also do not reflect naturalistic patterns of embodied visual experience. As a contribution to the small (but growing) number of multi-view object datasets that have been created to bridge this gap, we introduce a new video dataset called Toybox that contains egocentric (i.e., first-person perspective) videos of common household objects and toys being manually manipulated to undergo structured transformations, such as rotation, translation, and zooming. To illustrate potential uses of Toybox, we also present initial neural network experiments that examine 1) how training on different distributions of object instances and views affects recognition performance, and 2) how viewpoint-dependent object concepts are represented within the hidden layers of a trained network.
View on arXiv