gistfile1.txt

Questions to Discuss:

1) Use Hao Su’s paper or Sid’s observation on Multi-scale HoG Descriptor?
      a) How to use Hao Su’s paper
          i) Use rendered images and images 
          ii) Compare the distance on embedding space
      b) Multi-scale HoG performs very well
          i) How to quantatively measure the retrieval results
2) Current Bottleneck Part: Material CNN
      a) Fine-tune and generate a new model?
            i) Where is the training data?
                  MTurk?
                  Any small-size materialed shape dataset on Web?
                  Use current 150 shapes and Distort them to form more shapes?
      b) Use current model but add background to images?
            i) Like what Hao Su did in Joint Embedding paper
      c) How to detect small part like handles?
            i) Problem on the size of blobs used in Material CNN?
3) Improve HoG part:
      a) Image -> Rescale to 240x240, 120x120, 60x60, 30x30 -> Compute Edges -> Compute HoG -> Concatenate them
      b) Use Edge or Not, since the G in HoG is for gradient (Somewhat Edge Detection) -> Try all of them
      c) It seems a) is better!
4) Scraping Data from Amazon via EC2:
      a) Amazon keeps asking for OCR results
      b) Julian said he didn't encounter it but many encounters recently.
      c) Use Julian's data (Not for chairs but has a lot of chairs)
            i) Not too many
            ii) Quality is low
5) Should we start to use joint shape optimization?
6) Should we start to incorporate texture synthesis on shapes?