A Transformer-Based Object-Centric Approach for Date Estimation of Historical Photographs

The accurate estimation of the creation date of cultural heritage photographic assets is a challenging and complex task, typically requiring the expertise of qualified archivists, with significant implications for archival and preservation purposes. This paper introduces a new dataset for image date estimation, which complements existing datasets, thus creating a more balanced and realistic training set for deep learning models. On this dataset, we present a set of modern strong baselines that outperform previous state-of-the-art methods for this task. Additionally, we propose a novel approach that leverages “dating indicators” or “dating clues” through object detection and a self-attention based Transformer encoder. Our experiments demonstrate that the proposed approach has promising applicability in real scenarios and that incorporating “dating indicators” through object detection can improve the performance of image date estimation models. The dataset and code of our models are publicly available at https://github.com/cesc47/DEXPERT .