Event Location: NSH 1507
Bio: After spending his formative years in sunny Phoenix, AZ, Bryan received
his A.B. from Dartmouth College. He received his Ph.D. from MIT under
the supervision of William Freeman and Antonio Torralba. He is now an
INRIA post-doctoral fellow at Ecole Normale Superieure under Jean Ponce
and Andrew Zisserman. There, he pursues research in visual object
recognition and scene understanding. He is currently visiting CMU until
the end of August.

Abstract: We wish to build a large database of annotated images, spanning many
different scenes and objects, with absolute three-dimensional (3D)
coordinates. We infer the 3D coordinates given only the existing object
labels. For this, we learn “support” and “part-of” relationships that
hold between objects in a scene across the LabelMe dataset. In
addition, we reason about where objects contact their support and infer
the camera parameters using learned statistics of the absolute height of
objects. We show improved relative depth performance on the range data
of Saxena et al. captured over different street scenes. We have also
incorporated the 3D computation with the LabelMe annotation tool. We
show 3D outputs on images that were submitted through the annotation tool.