Visual tasks beyond categorization for training convolutional neural networks

Abstract

Humans can perceive a variety of visual properties of objects besides their category. In this paper, we explore- whether convolutional neural networks (CNNs) can also learn object-related variables. The models are trained for object position, size and pose, respectively, from synthetic images and tested on unseen held-out objects. First, we show that some object properties come "for free" from learning others, and pose-optimized model can generalize to both categorical and non-categorical variables. Second, we demonstrate that pre-training the model with pose facilitates learning object categories from both synthetic and realistic images.