rlm@66: #+title: Notes for "Special Topics in Computer Vision"
rlm@66: #+author: Robert McIntyre
rlm@66: #+email: rlm@mit.edu
rlm@66: #+description: 
rlm@66: #+keywords: 
rlm@66: #+SETUPFILE: ../../aurellem/org/setup.org
rlm@66: #+INCLUDE: ../../aurellem/org/level-0.org
rlm@66: #+babel: :mkdirp yes :noweb yes :exports both
rlm@66: 
rlm@66: * Fri Sep 27 2013
rlm@66: 
rlm@66:   Lambertian surfaces are a special type of Matt surface. They reflect
rlm@66:   light in all directions equally. They have only one parameter, the
rlm@66:   amount of energy that is absorbed/re-emitted.
rlm@66: 
rlm@66:   [[../images/adelson-checkerboard.jpg]]
rlm@66:   #+caption: Lol checkerboard illusion.
rlm@66: 
rlm@66:   Look into Helmholtz' stuff, it might be interesting. It was the
rlm@66:   foundation of both vision and audition research. Seems to have took
rlm@66:   a sort of Baysean approach to inferring how vision/audition works.
rlm@66:   
rlm@66:   - Homomorphic filtering :: Oppenhiem, Schafer, Stockham, 1968.  also
rlm@66:        look at Stockham, 1972.
rlm@66: 
rlm@66:   Edwin Land was Adelson's hero back in the day. He needed to create a
rlm@66:   color photo for the Polaroid camera. In order to process for
rlm@66:   automatic development of film, he had to get a good approximation for
rlm@66:   the illumination/reflectance decomposition that humans do, which he
rlm@66:   called Retinex.
rlm@66: 
rlm@66:   Cornsweet square wave grating is cool.
rlm@66: 
rlm@66:   - Retinex :: use derivatives to find illumination. Sort of
rlm@66:                implicitly deals with edges, etc. Can't deal with
rlm@66:                non-lambertian objects. 
rlm@66: 
rlm@66: 
rlm@66:   Adelson introduces the problem as an "inverse" problem, where you
rlm@66:   try to "undo" the 3-d projection of the world on your retina.
rlm@66: 
rlm@66:   On the functional view of vision : "What it takes" is to build a
rlm@66:   model of the world in your head. The bare minimum to get success in
rlm@66:   life is to have a model of the world. Even at the level of a single
rlm@66:   cell, I think you still benefit from models.
rlm@66: 
rlm@66:   Spatial propagation is ABSOLUTELY required to separate embossed
rlm@66:   stuff from "painted" stuff. Edges, likewise, MUST have spatial
rlm@66:   context to disambiguate. The filters we use to deal with edges must
rlm@66:   have larger spatial context to work, and the spatial extent of this
rlm@66:   context must be the ENTIRE visual field in some cases!
rlm@66: 
rlm@66: ------------------------------------------------------------
rlm@66: 
rlm@66: ** Illumination, shape, reflectance all at once
rlm@66:    
rlm@66:    What if we tried to infer everything together? Some images are so
rlm@66:    ambiguous it requires propagation from all three qualities to
rlm@66:    resolve the ambiguity.
rlm@66: 
rlm@66:    Brain has a competing painter, sculptor, and gaffer which each try
rlm@66:    to "build" the things in the world. There is a cost to everything
rlm@66:    such as paints, lights, and material, and then you try to optmize
rlm@66:    some cost function using these primitives.
rlm@66: 
rlm@66: 
rlm@67:    Horn, technical report, 1970
rlm@67: 
rlm@67: * Fri Oct  4 2013
rlm@67: 
rlm@67:   Student report. Talked about how you capture the appearance of a
rlm@67:   grape. It's actually quite compicated, involving gloss, spatial
rlm@67:   context, etc.
rlm@67: 
rlm@67:   Turbosquid seems interesting. They sell 3D models of stuff.
rlm@67: 
rlm@67:   BRDF -- bi-directional reflectance distribution function this shows
rlm@67:   how a surface will behave given lighting conditions. Lambertian is a
rlm@67:   simple parameterized instantiation of this.
rlm@67: 
rlm@67:   BSSRDF -- (SS = subsurface) 3D analogue of BRDF
rlm@67: 
rlm@67:   What would the 3D analiogue of texture be?
rlm@67: 
rlm@67:   (a : b : c) as (a + b + c : b + c : c) <-- this is just the golden
rlm@67:   ratio again!
rlm@67: 
rlm@67:   CURET BTF Database lol what's this
rlm@67: 
rlm@67:   This student went and gathered 1000 images of different large
rlm@67:   objects made of different materials. The images were gathered off of
rlm@67:   Flikr. 
rlm@67: 
rlm@84:   Then she gave another talk from someone else. It's about assigning
rlm@84:   materials to objects and then rendering them. The choice of
rlm@84:   materials is determined by some sort of expert system?
rlm@67: 
rlm@84:   They have made a neat looking interface for human entry of texture
rlm@84:   labeling or objects in scenes. The important elements were manual
rlm@84:   labels, dynamic display of the current selection, and undo.
rlm@84: 
rlm@84:   There are papers about mechanical turk engineering.
rlm@84: 
rlm@84:   lol http://opensurfaces.cs.cornell.edu/
rlm@84: 
rlm@84:   CUBAM is interesting 2010 "The multidimensional wisdom of crowds."
rlm@84:   "Neural Information processing systems" some sort of voting scheme.
rlm@84: 
rlm@84:   The point of this is apparently to do some kitchen makeover
rlm@84:   thing. You would take a picture of your kitchen, or you would look
rlm@84:   for a kitchen that looks like yours, and then you would be able to
rlm@84:   investigate different textures for your own kitchen.
rlm@84: 
rlm@84:   apparently "Label ME" has never been appropriately
rlm@84:   crowdsourced. Turns out that you get better work if you don't use
rlm@84:   idiots to do the work lol.