Mini-Resizer: A Minimalist Learnable Resizer for Image Geolocation

Abstract:
For practical application scenarios in the field of image geolocation, it is often necessary to resize images into a uniform size to fit deep learning training tools. However, the traditional resizing process would cause information loss, thus leading to limited geolocation performance. To solve this problem, this paper proposes a new learnable resizer, that maps the features to a high-dimensional space through convolution, then performs feature filtering to obtain a resized image, which is more suitable for deep learning-based geolocation tasks rather than the human eye. Compared with the existing methods, the proposed method has better geolocation performance on the Piitburgh30K dataset.