shape-cmp

shape similarity experiment

• Followed this tutorial: https://blog.mturk.com/tutorial-how-to-label-thousands-of-images-using-the-crowd-bea164ccbefc

• Overall strategy is just to pre-compile all the images, and then present users with a binary choice (Left or Right).

• Mturk sandbox: https://requestersandbox.mturk.com/create/projects -- have to create in there and then login as a worker in workersandbox -- search for your expt *not by name but by keywords!"

• Mturk login: https://requester.mturk.com/create/projects

• AWS S3 account used to host the images: https://console.aws.amazon.com/billing/home?#/account

  • Instructions for making things public has changed since the tutorial -- I ended up just disabling all the various public ACL kinds of options under Permissions, and then it allowed me to make them public

  • After uploading new files (overwrote existing) I had to go up to higher dir and use context menu, "make public" to get them to be readable -- if it doesn't show you images during development in sandbox etc, then the files are not readable!

• shape-cmp.go generates the experimental trials, compositing images previously generated using the imgproc.go code, which processes image files captured from model, avail for download here: https://grey.colorado.edu/downloads/wwi_emer_imgs_20fg_8tick_rot1.tar.gz

  • Current version: runs image through V1 filters and then inverts back out using random selection of max pool unit -- fairly unrecognizable but definitely gives an overall impression of shape..
  • Previous: turns all the blue background color to white, and everything else to black, and then blurs the resulting image with a bild blur.Box convolution filter with a radius of 25. Produces pretty blurry images.