Misidentified is a practical critique of culturally biased datasets. It raises the question of machine perception of different cultures. These datasets are in analogy the modern archeological record of material culture, of which we perceive—contemporary— societies through.

This question was sparked by when ML model, im2text, identified— or misidentified— a full covered woman with her son as: "A man sitting on a curb holding his phone".

In this project, we are testing several images through im2txt to generate rather irrelevant captions to use them as gallery labels with the imagery artifacts. The selection of images is based on assigned factors for the purposes of the Artathon: One, candid culturally relevant images and two, mostly men to highlight the "Shimagh" element for its aesthetics.

However, one image has stood out in the results and inspired even more experimentation with it. That is the "a bunch of vases that are sitting on a table."; which is in fact a group of Saudi men wearing their traditional red and white Shimagh in an auditorium. We took the image through a minor session in Deep Dream, taking the result and Style Transferring it with a mask of flowers we made in photoshop; their synthesis was taken again through Deep Dream but for multiple sessions to finally create the masterpiece of the image gallery: A Garden of Men.

The images are exhibited as a photo gallery with captions and translation into Arabic. It is designed to gradually reveal the layers of misidentification. The user experience starts with full view of familiar photos and images until he or she reads through the labels and its translation. Only then, the title of the exhibition/project and the witnessed results come into a full circle.

This being said, Misidentified is a pseudoarchaeological project which aspires to afford diverse AI-generated records to encourage and open the conversation about our future archive.

Share this project:
×

Updates