On the Influence of Viewpoint Change for Metric Learning

Physical objects imaged through a camera change their visual representation based on various factors, c.g., illumination, occlusion, or viewpoint changes. Thus, it is the inevitable goal in computer vision systems to use mathematical representations of these objects robust to various changes and yet sufficient to determine even minor differences to distinguish objects. However, finding these powerful representations is challenging if the amount of data is limited, such as in few-shot learning problems. In this work, we investigate the influence of viewpoint changes in modern recognition systems in the context of metric learning problems, in which fine-grained differences differentiate objects based on their learned numeric representation. Our results demonstrate that restricting the degrees of freedom, especially by fixing the virtual viewpoint using synthetic frontal views, elevates the overall performance. We await that our observation of an increased performance using rectified patches is persistent and reproducible in other scenarios.