Outline • Theoretical approaches to computer vision – Visual perception as information processing • Problems in Computer Vision – – – – Classification Segmentation Recognition Motion analysis Visual Perception as an Inverse Problem • Retinal images are generated by the light reflected from the 3-D world – The image formation is determined by the laws of optics – The area of image rendering is called computer graphics • Vision as an inverse problem – Get from optical images of scenes back to knowledge of the objects that gave rise to them 5/29/2016 Visual Perception Modeling 2 Vision as a Heuristic Process • Visual system makes a lot of assumptions about the nature of the environment and conditions under which it is viewed – These assumptions constrain the inverse problem enough to make it solvable most of the time – The resulting solution will be veridical if the assumptions are true – Vision is a heuristic process in which inferences are made about the most likely environmental condition that could have produced a given image 5/29/2016 Visual Perception Modeling 3 Perception as Bayesian Inference • Images I are observations • Scene properties S are not known • p(S) specifies the prior knowledge about the scene – The knowledge you have without looking at the image • Bayes’ rule p( I | S ) p( S ) p( S | I ) p( I ) p( S | I ) p( I | S ) p( S ) 5/29/2016 Visual Perception Modeling 4 Four Stages of Visual Processing • • • • Image-based stage Surface-based stage Object-based stage Category-based stage 5/29/2016 Visual Perception Modeling 5 Image-based Stages • Most theorists agree that initial stage is not the only representation based on a twodimensional retinal organization – It includes image-processing operations • Local edge and line detection • Region detection • Correspondence between left and right eyes – Marr called this representation primal sketches • Raw primal sketch • Full primal sketch 5/29/2016 Visual Perception Modeling 6 Representation in Early Vision • Local spatial/frequency representation – The representation should be • Local • Orientation-tuned • Frequency-tuned – Gabor filters – Wavelet transformation • Image compression 5/29/2016 Visual Perception Modeling 7 Gabor Filters 5/29/2016 Visual Perception Modeling 8 Surface-based Stage • Recovery of intrinsic properties of visible surfaces • Surface layout – The spatial distribution of visible surfaces within the 3-D environment • Explicit surface-based representation – 2.5-D sketch – Intrinsic images • Intrinsic properties to surfaces 5/29/2016 Visual Perception Modeling 9 Surface-Based Stage – cont. • Surface primitives – Local patches of 2-D surface within a 3-D space • Three-dimensional geometry – Projective geometry • Viewer-centered reference frame 5/29/2016 Visual Perception Modeling 10 Surface-Based Stage – cont. • Cues for surface representation – – – – Stereopsis Motion parallax Shading and shadows Pictorial properties • • • • 5/29/2016 Texture Size Shape Occlusion Visual Perception Modeling 11 Object-Based Stage • Some form of true 3-D representation – Includes unseen and occluded surfaces • Explicit representations of whole objects • Two ways of constructing object representation – Extend the surface-based representation – Infer 3-D objects from 2-D images 5/29/2016 Visual Perception Modeling 12 Object-Based Stage – cont. • Volumetric primitives – Descriptions of truly 3-D volumes • Three-dimensional geometry – Geometry in 3-D space • Object-based reference frame – Spatial relations among the volumetric primitives are represented by intrinsic structures among volumetric structures 5/29/2016 Visual Perception Modeling 13 Category-Based Stage • Final stage concerns with recovering fully the functional properties of objects – Functional properties through categorization – Properties directly from visible characteristics 5/29/2016 Visual Perception Modeling 14 Top-down vs. Bottom-up Processes • Bottom-up processing – Data driven processing – Take a lower-level representation as input and create or modify a higher-level representation • Top-down processing – Expectation-driven processing – Processes that take a higher-level representation as input and produce or modify a lower-level representation 5/29/2016 Visual Perception Modeling 15 Neural Network Approaches • Neural networks are based on the assumptions that human vision depends heavily on the massively parallel structure of neural circuits in the brain – Multiple Layer Perceptrons • Input layer • Hidden layer • Output layer 5/29/2016 Visual Perception Modeling 16 Problems in Computer Vision • Given a matrix of numbers representing an image, or a sequence of images, how to generate a perceptually meaningful description of the matrix? – An image can be a color image, gray level image, or other format such as remote sensing images – A two-dimensional matrix represents a single image – A three-dimensional matrix represents a sequence of images • A video sequence is a 3-D matrix • A movie is also a 3-D matrix 5/29/2016 Visual Perception Modeling 17 Image Classification • Given some types through examples, identify the type of a new image 5/29/2016 Visual Perception Modeling 18 Image Segmentation • Partition the images into homogenous regions – Widely studied problem – A very difficult problem – An important problem A texture image 5/29/2016 Visual Perception Modeling 19 Object Recognition • Object recognition – Recognize objects in a constrained environment – Identify objects from images A cheetah image 5/29/2016 Visual Perception Modeling 20 Video Sequence Analysis • Motion analysis – Compute motion from images – Motion segmentation • Video sequence analysis – Derive models automatically – Enhanced TV 5/29/2016 Visual Perception Modeling 21