Learning Scale-Invariant Object Representations with a Single-Shot Convolutional Generative Model

Piotr Zielinski, Tomasz Kajdanowicz

Abstract: Contemporary machine learning literature highlights learning object-centric image representations’ benefits, i.e. interpretability, and the improved generalization performance. In the current work, we develop a neural network architecture that effectively addresses the task of multi-object representation learning in scenes containing multiple objects of varying types and sizes. In particular, we combine SPAIR and SPACE ideas, which do not scale well to such complex images, and blend them with recent developments in single-shot object detection. The method overcomes the limitations of fixed-scale glimpses’ processing by learning representations using a feature pyramid-based approach, allowing more feasible parallelization than all other state-of-the-art methods. Moreover, the method can focus on learning representations of only a selected subset of types of objects coexisting in scenes. Through a series of experiments, we demonstrate the superior performance of our architecture over SPAIR and SPACE, especially in terms of latent representation and inferring on images with objects of varying sizes.