#+title: Notes for "Special Topics in Computer Vision"

 #+author: Robert McIntyre

 #+email: rlm@mit.edu

 #+description: 

 #+keywords: 

 #+SETUPFILE: ../../aurellem/org/setup.org

 #+INCLUDE: ../../aurellem/org/level-0.org

 #+babel: :mkdirp yes :noweb yes :exports both

 

 * Fri Sep 27 2013

 

   Lambertian surfaces are a special type of Matt surface. They reflect

   light in all directions equally. They have only one parameter, the

   amount of energy that is absorbed/re-emitted.

 

   [[../images/adelson-checkerboard.jpg]]

   #+caption: Lol checkerboard illusion.

 

   Look into Helmholtz' stuff, it might be interesting. It was the

   foundation of both vision and audition research. Seems to have took

   a sort of Baysean approach to inferring how vision/audition works.

   

   - Homomorphic filtering :: Oppenhiem, Schafer, Stockham, 1968.  also

        look at Stockham, 1972.

 

   Edwin Land was Adelson's hero back in the day. He needed to create a

   color photo for the Polaroid camera. In order to process for

   automatic development of film, he had to get a good approximation for

   the illumination/reflectance decomposition that humans do, which he

   called Retinex.

 

   Cornsweet square wave grating is cool.

 

   - Retinex :: use derivatives to find illumination. Sort of

                implicitly deals with edges, etc. Can't deal with

                non-lambertian objects. 

 

 

   Adelson introduces the problem as an "inverse" problem, where you

   try to "undo" the 3-d projection of the world on your retina.

 

   On the functional view of vision : "What it takes" is to build a

   model of the world in your head. The bare minimum to get success in

   life is to have a model of the world. Even at the level of a single

   cell, I think you still benefit from models.

 

   Spatial propagation is ABSOLUTELY required to separate embossed

   stuff from "painted" stuff. Edges, likewise, MUST have spatial

   context to disambiguate. The filters we use to deal with edges must

   have larger spatial context to work, and the spatial extent of this

   context must be the ENTIRE visual field in some cases!

 

 ------------------------------------------------------------

 

 ** Illumination, shape, reflectance all at once

    

    What if we tried to infer everything together? Some images are so

    ambiguous it requires propagation from all three qualities to

    resolve the ambiguity.

 

    Brain has a competing painter, sculptor, and gaffer which each try

    to "build" the things in the world. There is a cost to everything

    such as paints, lights, and material, and then you try to optmize

    some cost function using these primitives.

 

 

    Horn, technical report, 1970

 

 * Fri Oct  4 2013

 

   Student report. Talked about how you capture the appearance of a

   grape. It's actually quite compicated, involving gloss, spatial

   context, etc.

 

   Turbosquid seems interesting. They sell 3D models of stuff.

 

   BRDF -- bi-directional reflectance distribution function this shows

   how a surface will behave given lighting conditions. Lambertian is a

   simple parameterized instantiation of this.

 

   BSSRDF -- (SS = subsurface) 3D analogue of BRDF

 

   What would the 3D analiogue of texture be?

 

   (a : b : c) as (a + b + c : b + c : c) <-- this is just the golden

   ratio again!

 

   CURET BTF Database lol what's this

 

   This student went and gathered 1000 images of different large

   objects made of different materials. The images were gathered off of

   Flikr. 

 

   Then she gave another talk from someone else. It's about assigning

   materials to objects and then rendering them. The choice of

   materials is determined by some sort of expert system?

 

   They have made a neat looking interface for human entry of texture

   labeling or objects in scenes. The important elements were manual

   labels, dynamic display of the current selection, and undo.

 

   There are papers about mechanical turk engineering.

 

   lol http://opensurfaces.cs.cornell.edu/

 

   CUBAM is interesting 2010 "The multidimensional wisdom of crowds."

   "Neural Information processing systems" some sort of voting scheme.

 

   The point of this is apparently to do some kitchen makeover

   thing. You would take a picture of your kitchen, or you would look

   for a kitchen that looks like yours, and then you would be able to

   investigate different textures for your own kitchen.

 

   apparently "Label ME" has never been appropriately

   crowdsourced. Turns out that you get better work if you don't use

   idiots to do the work lol.