About
As artificial intelligence progresses, so-called 'convolutional networks' (ConvNets) now exist that can recognise objects within photographic images. Constant Dullaart has retrained these image recognition networks to include European artefacts, creating an image dataset. He asks, 'How can Europe's diverse cultural output be represented within this dataset, and what is the networks' capacity to recognise what is deemed European?' By illustrating the networks' ability to draw out cultural bias, Dullaart shows mechanised image interpretation's understanding of Europe in 2017.
Constant Dullaart was born in the Netherlands. He reflects on the cultural and social effects of communication and image processing technologies. His work includes websites, performances, fake armies and manipulated found images, presented both offline and in the public space of the Internet.
ConvNet Training
We used Caffe transfer learning and the BVLC Caffenet architecture to train the model. The EuroNet classes were augmented with about 150 random classes from the original BVLC Caffenet to provide more interclass separation. These images were obtained from the original Imagenet project
The network was trained on a GTX 1080 GPU for about 12 hours. The images were then rendered using the Plug and Play Generative Networks by Anh Nguyen.
You can download the EuroNet caffemodel and generate your own Chancellor Angela Merkel, Vodka, or Weiner Schnitzel.
Downloads
Credits
ImageNet.xyz was inspired by the original Imagenet, a collection of 14 million images compiled by Stanford and Princeton University. Like the original Imagenet, we only provide URLs to bypass copyright issues.
The following tools and frameworks were used in the development of the ImageNet.xyz EuroNet dataset:
- Plug and Play Generative Networks by Anh Nguyen
- Caffe: a fast open framework for deep learning
- imsearch-tools by Ken Chatfield
ImageNet.xyz concept developed in collaboration with, and coded by Adam Harvey