Stanford Vision Reading Group
<p class="MsoNormal" >CHANG, Angel, Text Description to 3d Scene Geneeration, Feb 3, 2014</p>
<p>
in Fei-Fei's computer vision course; in Gates 120 at 2:15PM.</p>
<p>
eg a cat looks out the window.... want to generate model from that...</p>
<p>
input phrase goes into constraint parsing and that goes into scene inference engine.?</p>
<p>
need to know the category of objects</p>
<p>
eg object is chair with parts = back, seat, legs
attributes wood, yellow
semantic faces: right, left front of chri.</p>
<p>
next need to manipulate 3d model attributes...
say change seat of chair to have green cloth;
(a lot of knowledge is involved)</p>
<p>
see Coyne and Sproat in 2001 WordsEye
a zebra is standing 20 feet in front of a barn with sky in background.</p>
<p>
uses Viget which is an extension of FrameNet... see WordsEye;</p>
<p>
VigNet is an extension of FrameNet (? from Columbia... from Bob Coyne)
it integrates world know. and graphical know.</p>
<p>
the data in it is collected by manual annotation and corwdsourcing;
eg washer is infront of the sink.</p>
<p>
she shows a simple, early demo:</p>
<p>
she types in "there is a red chair and three blue bottles"
it generates those objects correctly...</p>
<p>
feifei... any stochastic aspect to this ?</p>
<p>
what does to the RIGHT of the chair mean (do u mean egocentric or object centric)</p>
<p>
I ask abt geomtric modeling (eg where is the floor)...
A from Manolis: inferred from pre-existing scenes that were created using modeling software.</p>
<p>
*88
Manolis continues this seminar now...</p>
<p>
why do want the 3d scenes;</p>
<p>
we want to learn affordances...</p>
<p>
important distinction btwn static and synamic scenes;</p>
<p>
more general case: eg watch TV (and seeing a couch in front of the TV)</p>
<p>
where can a human sit... in the chair and on the couch and facing the TV;</p>
<p>
can ask people for action scenes...
eg cut an onion.. and a person creates kitchen counter with onion on a cutting board with a knife at the side.</p>
<p>
or "play the piano" user creates aperson seated at a piano.</p>
<p>
turn to real world and have video cameras observing people as they engage in actions...</p>
<p>
Goal is to learn agent actions... a controller for a virtual agent</p>
<p>
eg learn the staes and probs... sit in couch, streatch arms</p>
<p>
Jeff Zachs is working on this; a student of Barabara Tversky... now a prof...</p>
<p>
want parts decomposition, appearance, affordances, materials;
attributes of objects (eg materials), semantic faces (front, r, l);</p>
<p>
</p>