Talk:Image Colourization using Deep Learning

From UBC Wiki
Jump to: navigation, search


Thread titleRepliesLast modified
Critique021:49, 12 March 2018
Some points to help your article019:53, 12 March 2018
Critique017:06, 12 March 2018

This page doesn't look quite finished yet. Here's my comments for what they're worth:

  • Let us know what LAB is short for
  • Some categorization is going on in the training labels for both papers (it seems like this categorization is the contribution of the first paper), however it was not clear what or how
  • Self organizing map deserves more discussion, as this is the improvement compared to the first paper if I understand correctly
  • Personally I'd like to see some discussion on the results, instead of just some images thrown out there and then abrupt stop to the section
21:48, 12 March 2018

Some points to help your article

Hi Amin, Here are a few suggestions that I feel might help your page:

  • Change the “Title” itself to “Colourizing grayscale images using deep learning and neural network”. You could do this by editing the topmost edit.
  • Fill up the Abstract, “Builds On” and “Related To” page. Maybe, builds on could be the wiki link to Deep Learning, Neural Network, CNN. Your introduction section could fill up the Abstract section. It could explain in brief about the two papers of your choice.
  • You could reword this: “Choosing reasonable colors for the computer is a challenging task” as “The same task of choosing colors is very challenging for a computer”.
  • How many channels are there in CEILAB? What is their significance?
  • In the section “Objective Function”, what does R indicate? What is the intuition of Y? You could clearly mention what the two Y’s are when you calculate the loss. What is the loss used for? Why is the colorization problem multimodal? What does a grid of space mean in terms of an image? What does Q=312 indicate? What is Q? What does “in-gamut” mean? How does Regression, classification, classification with rebal mean? How is ground truth identified? Is ground truth used in training from some database?
  • It is unclear what your two papers were. What is the incremental addition/problem that was solved by paper 2.


19:53, 12 March 2018

Critique from Gudbrand:

Interesting topic and a good start. I don't think you're quite finished, but I'll give you my input now anyway.. - You're missing the abstract/related work/builds on section. - The sentence "generating a hallucinating..." doesn't parse. Do you mean "or"? - "Graphic experts" -> "Graphics experts"? - "easier for human" -> "easier for humans?" - I believe the term "image colourization" should not be capitalized. - You introduce several colour spaces without really explaining what they are. Perhaps have blue links to them? - To me it's not exactly clear what CIELAB/CIE really is. It seems it's a colour space with some nice properties, but how does it work? - You keep referring to "these works" or "this paper" without having mentioned that you are reviewing papers. - Can you explain what the "a" and "b" channels are? - What you call "architecture 1" should really be called "figure 1", no? - You mention CNN without explaining what it is. Ok, so everyone in the course knows what you mean, but perhaps use a blue link or write out the full term "convolutional ... " - You completely lost me at the "Q=313/in-gamut" part. Do not expect that your reader has read the paper! - Lightness and Light, etc. should not be capitalized - Perhaps make more explicit the comparison/deltas between the two papers? Is it just SOMs or what?

Generally I'd like to see some more explanatory details. Also more references and links, if it can be useful.

17:06, 12 March 2018