Deep Training on History

Automatic Annotation And Visualization Of Archival Imagery Databases

The impressive performance of modern computer vision comes at the price of large number of images that are annotated by humans. It takes millions of images that are labelled with hundred thousand of object classes to teach a computer to “see” the world. Notably, the collected image sets for visual learning are all from the contemporary era, thus our “intelligent” companions have not seen our past yet. To teach the computers our history, we need to show them how our world looked like before the advent of digital cameras or else our computer assistants are unable to recognize and detect objects and semantics of past.

Currently trained computer models, where used in the context of visual history, often take the image style, e.g., being black and white or blurred, as a discriminative attribute rather than the image content. In this project, we spot the invaluable amount of visual archival data available at KB for training style-agnostic computer models that can detect semantic objects regardless of the image representation.

The outcome of DepTH project contributes to the automatic annotation and visualization of historical images and extending the linked data between historical and contemporary visual data repositories. The trained computer models may also be released to the public use, which is of value for anyone who is bridging historical and contemporary visual record and thus a reusable and sustainable research project.