Low-shot learning for object recognition, detection, and segmentation

Abstract

Deep Neural Networks are powerful at solving classification problems in computer vision. However, learning classifiers with these models requires a large amount of labeled training data, and recent approaches have struggled to adapt to new classes in a data-efficient manner. On the other hand, the human brain is capable of utilizing already known knowledge in order to learn new concepts with fewer examples and less supervision. Many meta-learning algorithms have been proposed to fill this gap but they come with their practical and theoretical limitations. We review the well-known bi-level optimization as a general framework for few-shot learning and hyperparameter optimization and discuss the practical limitations of computing the full gradient. We provide theoretical guarantees for the convergence of the bi-level optimization using the approximated gradients computed by the truncated back-propagation. In the next step, we propose an empirical method for few-shot semantic segmentation: instead of solving the inner optimization, we propose to directly estimate its result by a general function approximator. Finally, we will discuss extensions of this work with the focus on weakly-supervised object detection when full supervision is not available for the few training examples.