Vision is versatile. It lets us navigate our environment, grasp and manipulate tools, catch or avoid projectiles, and recognize familiar objects and people. This versatility stems from a low level of visual processing that somehow produces, from retinal measurements, useful intermediate representations of depth, surface orientation, motion, and other intrinsic scene properties.

During the past ten years, my research group has been using various tools---from applied mathematics and signal processing, to computer graphics and computational imaging---to help build a catalogue of the ways in which local patterns of image brightness can constrain the intrinsic scene properties within a small receptive field. But suppose our catalogue of these local constraints is complete. How should we apply and combine them for vision?

I will try to answer this question by introducing a computational framework for merging diverse, noisy scene constraints at many locations and scales across the visual field. It can be described as a large collection of distributed computational units, one for each receptive field of a certain position and size. Regardless of its associated position and size, each of the units iteratively performs the same set of simple calculations, and it iteratively trades messages with other units through sparse connections between scales. I will show results of applying this framework to estimate depth from stereo images, and to recover surface normals from monocular diffuse shading.

The talk is based on two papers:

1. Ayan Chakrabarti, Ying Xiong, Steven J. Gortler, and Todd Zickler, "Low-level Vision by Consensus in a Spatial Hierarchy of Regions." CVPR 2015. http://ttic.uchicago.edu/~ayanc/consensus/
2. Ying Xiong, Ayan Chakrabarti, Ronen Basri, Steven J. Gortler, David W. Jacobs, and Todd Zickler, "From Shading to Local Shape." PAMI 2015. http://vision.seas.harvard.edu/qsfs/

Todd Zickler is visiting Victoria University for six months, January to July 2017, based in the School of Engineering and Computer Science.