Intro

We randomly sampled 1000 images out of the ~138k that are currently accessioned into SDR and generated labels using the Google Vision API [GV]1 and the Clarifai Predict API [CP].

Code to generate GV labels is here: https://github.com/sul-dlss-labs/google-vision-ai

Catal project is here: http://catalhoyuk.com

SUL AI studio: https://sites.google.com/stanford.edu/sul-ai-studio/

SUL AI studio talk is here.

For various reasons we had some attrition and ended up with 766 images. Results shown here are based on those.

0.1 Comments, ideas in random order

  • It could be useful for discoverability if the images are embedded in a much larger dataset, but not sure about within the image dataset itself.

  • Instead of looking for positive identification turn it around and look for things that are NOT excavation or artifacts. For example, ‘buildig’ gets us images of the shelter, so allows us to weed out relevant images(?).

  • Combine several labels to separate people (man, woman, person, group, …) from buildings (soil, wall, dirt, …) from X-finds etc.

  • Use GV_OCR to identify images with whiteboards, though the rendered text is not very usable.

  1. Many thanks to Peter Mangiafico.↩