Skip to content

Week 3 Meeting Notes

Jump to bottom
Richard Luo edited this page Feb 24, 2021 · 2 revisions

Notes

  • We found an auto Canny filter that properly sets the proper max and min val based on the input image. Even with this, it sometimes creates poor edge detection on landscapes and the presence of water
  • After running Canny, we can do cv2.bitwise_not to invert the black and white pixels, so it looks more like a traditional drawing
  • The dataset from last time is mostly trees, but we will use the 2.9 GB set as our starting place
  • This dataset does not really produce realistic drawings for a lot of the images, so we may need to find a better topic other than landscapes
  • Everyone should download this dataset and then create their own separate developer branch to experiment with code locally
  • We could potentially use CycleGAN and Pix2Pix as they already have the architecture set up for this project
  • After creating the edge detected version of each image, we will pair them with the corresponding original image using a simple Python script

To Do

  • Send a message to Neil about which architectures we should use
  • Everyone should make their own GitHub dev branch
  • Find a dataset that produces realistic edge detection or already has drawings included (animated characters seem promising, probably because they have defined shapes and edges)

Promising Datasets

Clone this wiki locally