Challenges to Computer Vision Larry Davis Computer Science

Challenges to Computer Vision Larry Davis Computer Science Department Institute for Advanced Computer Studies University of Maryland College Park, MD 20742

Obstacles to human-level machine vision Segmentation – Finding what to see n Object representation – What things look like n Visual learning – Of the What and How n Interface to cognition – Reasoning about what is seen n

Segmentation n Identification of “significant” groups of pixels – edge and local feature detection – edge and local feature grouping – region analysis texture n color n – motion analysis From Adelson, MIT

Segmentation of moving objects against “fixed” backgrounds

Space shuttle example

Segmentation challenges n Integration – Scales in space and time – Cues n Attention

Representing 3 -D objects n Object versus viewer centered representations – shape + texture

Appearance based matching example (Columbia, msoft)

Outdoor example – traffic signs

Pedestrian detection results

Representing 3 -D objects n Articulated and deformable objects

Object-centered deformable objects n Metaxas and Terzopoulos

Representing 3 -D objects n Object versus viewer centered representations – shape + texture Articulated and deformable objects n General object classes n

Representing 3 -D objects n Object versus viewer centered representations – shape + texture Articulated and deformable objects n General object classes n Form and function n

Visual Learning n Learning representations – recovery – coding n Learning control information – process parameters – rules of engagement

Recovery of 3 -D models Stereo n Structure from Motion n Prior shape models as constraints on recovery n

Multi-view structure recovery (Torr)

Interface to cognition n Knowledge based vision systems