RepMet: Representative-based metric learning for classification and one-shot object detection

• Main Authors: Karlinsky, Leonid, Shtok, Joseph, Harary, Sivan, Schwartz, Eli, Aides, Amit, Feris, Rogerio, Giryes, Raja, Bronstein, Alex M.
• Format: Report
• Language: unknown
• Published: arXiv 2018
• Subjects: Computer Vision and Pattern Recognition cs.CV; FOS Computer and information sciences; DML
• Online Access: https://dx.doi.org/10.48550/arxiv.1806.04728; https://arxiv.org/abs/1806.04728

Distance metric learning (DML) has been successfully applied to object classification, both in the standard regime of rich training data and in the few-shot scenario, where each category is represented by only a few examples. In this work, we propose a new method for DML that simultaneously learns the backbone network parameters, the embedding space, and the multi-modal distribution of each of the training categories in that space, in a single end-to-end training process. Our approach outperforms state-of-the-art methods for DML-based object classification on a variety of standard fine-grained datasets. Furthermore, we demonstrate the effectiveness of our approach on the problem of few-shot object detection, by incorporating the proposed DML architecture as a classification head into a standard object detection model. We achieve the best results on the ImageNet-LOC dataset compared to strong baselines, when only a few training examples are available. We also offer the community a new episodic benchmark based on the ImageNet dataset for the few-shot object detection task.