Object Recognition by Stochastic Metric Learning

Abstract

Descriptors extracted from deep neural networks have been shown to be very discriminative, for example networks such as those trained on the large, very general ImageNet dataset have been used to extract descriptors robust for a variety of image classification tasks. Such retrieval systems utilize feature locality, for example Approximate Nearest Neighbour. Our goal is to use such descriptors as part of a large scale object instance recognition and retrieval system. We propose using deep nonlinear metric learning on Convolutional Neural Networks to learn features with good locality. In particular we worked with two related methods, Neighborhood Components Analysis (NCA) and the related Mean square Error’s Gradient Minimization (MEGM).

We utilize a nonlinear form of MEGM as an alternative to NCA and propose some stochastic sampling methods to apply these (normally batch) methods to larger datasets with minibatch Stochastic Gradient Descent (SGD). On a larger scale we found the methods difficult to train, failing to converge or generalizing very badly depending on training method or parameters. This led us to go back to a smaller dataset and examine the factors which lead to good generalization with this form of training.

We found on a small subset of the RGB-D dataset, surprisingly stochastic sampling methods generalized much better with small batch sizes, which acted as a form of regularization. When trained with larger batches, or as a full batch, the dataset was overfit. Given the correct parameters, descriptors extracted performed well at the Nearest Neighbour task and exceeded the performance of those extracted by applying standard supervised training.