I took the first 100 images from a 1% random sample of reference images in Calisphere. These were limited to images larger than 150 x 150 in the thumbnail server (min size for the google vision API). I resized them to be no larger than 640 x 480 and uploaded them to google cloud storage. Then, I wrote a quick python script to run all the images through google vision API and save the reuslts.
https://cloud.google.com/vision/
I did this by getting a 60 day free trial account with a $300 test credit. If I'm reading the prices correctly then it would cost ~$12,000 to run all the google vision API features on all ~630k Calisphere images, or $20 per 1,000 images.
The following features were requested from each file:
- LANDMARK_DETECTION
- LABEL_DETECTION
- SAFE_SEARCH_DETECTION
- TEXT_DETECTION
- FACE_DETECTION
- LOGO_DETECTION
The face dection and text detection were run on the 640x480 image. Recommended sizes for face detection is 1600x1200 and for text it is 1024x768. A next test might be to re-run these feature on higher resolution versions (where available, these samples are from the "thumbnail" server).