Identifying locations in the Bingley Photographic Collection with AI

Before everything starts, how do you think Digital Preservation works?

Cultural Digital Preservation is always an interesting topic that has piqued my attention. I am glad that the University of Leeds offered me this internship to work on something I am quite excited about. This blog post aims to share some of the progress that has been made so far and some of the interesting findings during research and experiments!

Let's briefly talk about my background. I am from Hong Kong, where I worked as a STEM instructor and teaching in schools Intangible Cultural Heritage with Minecraft, as well as different STEM courses such as Coding, SwiftPlayground etc... But I mainly worked with using Minecraft to teach Hong Kong Intangible Cultural Heritage, and it works as a way of digital preservation as well! Apart from this computing background, I am currently a music student in University of Leeds, also a photography enthusiast. After being given this role, I spent some time to figure out which set of Special Collections I should interact with (there are loads of special collections in University of Leeds Library, which are collections that are unique, only in our library). I narrowed it down to only two collections I wanted to work on, which are the Godfrey Bingley Photographic Archive and the Chinese Collection. Unfortunately, it was way too much work to do both of them, and so I need to pick one and there is it, the Bingley Collection!

The Bingley Collection was donated to the University of Leeds Library in 1913 by Godfrey Bingley, a Leeds-born photographer, who donated more than a thousand of his own images. My project set out  to use the photographs as data, so I began by brainstorming a set of innovative AI-powered experiments. My first ever idea: Geoguessr trivia.

Geoguessr is a place-guessing game where users need to pin a location they think the view is on the map; the closer, the higher the marks. And before setting up a quiz, we have to accurately check the coordinates of the photo. It is partly inspired by Timeguessr, but due to the lack of a custom version, we can only create a custom quiz on Geoguessr.

The image above shows some of the data and unfortunately, the map coordinates are not accurately pinning the location on the map. So what tools we can use?

This leads me to another idea, Bot Training.

However, a few different platforms provide self-training bots, and after a few initializations, I decided on ChatGPT. It may not be fully accurate, but I found out that due to the configuration setup, it targets black and white (B&W) photography. It turns out to be less accurate with colorized photos.

Apart from GPT engineering, I tried a more originally structured AI setup using Huggingface.co. First of all, I needed to upload and fetch my dataset to the platform. However, it got stuck before the fetching of photos. The debugging did not go well. I used ChatGPT for code-writing and debugging. The error was not solved until I changed the platform to another AI assistant platform, Claude.ai. It quickly uploaded the dataset to Huggingface, and the code worked so well!

This is the first ever completed code-bot training, with an accuracy rate of 30.2%.

However, it is understandable that just from a picture it is hard to guess, therefore here is a little test for you: see if you can guess the location of the picture below!

According to the picture features, can you guess where this is, and why? Before you scroll down and look at the answer, let's see how the bot tried to work it out...

Using this analysis, the bot quickly provide an answer.

However, we are focussed on Yorkshire, and Peterborough is not in Yorkshire, so here we can give the bot a little clue and see how that change the output:

And YES! It is in Ripon. And this basically gives you an idea how an AI usually analyses a picture and how further prompts can make the result more accurate.

Now let's try another image:

Where do you think this is, based on the image features, and why? Again, let's see how the bot did before scrolling down to find the answer...

Did you and the bot reach the same conclusion?

And yes! It is Fountains Abbey.

This is the work I have done so far, and I am so glad that I got these opportunities to work with these amazing people at DCCH, as well as attending conferences about digital heritage. Also, this is my first time working in an English-speaking atmosphere. It confused me a bit at first, but so far, I feel comfortable and enjoy working. I hope to see more progress from myself, although I need to take care of my studies as well. In spite of the challenges, I have found the experience to be incredibly rewarding. The exchange of ideas and perspectives in a multicultural setting has broadened my understanding and enriched my development apart from music education. I look forward to balancing my work and academic responsibilities to continue growing in both areas.

Samuel Yeung