Lecture 3 Flashcards
Discuss the challenges of object perception
Different classes of things recognized - from different angles and when partly occluded. Occurs rapidly and without error. It appears effortless but is very complex. Objects overlap and vary enormously in their visual properties but we see them as the same thing. We accurately recognise objects over a wide range of viewing distances and orientations
Define perception
The ability of humans to recognise familiar, concrete things such as items of furniture, vechiles, fruits and vegetables.
Discuss the differences between human perception and computers
People take in information at multiple scales - registering large-scale shapes and patterns as quickly as small details. Whereas computers start with pixels in images and build up. We usually see the global features first, but we do have some control over which we see first.
What much be accomplished to recognise an object
See the basic features in the visual scene. Perceive organisation in the features. Perceive shape. Compare the shape percept to memory for known shapes. Make a decision about whether the object is familiar, and tap into knowledge about the object.
What is simultanagnosia
The inability to perceive how parts fit together. For example, in a picture they could pick out curtains or a child but would be unable to perceive the whole scene
Discuss the process of perceiving shape
Figure vs Group. Impossible fork - this illusion works because we process the global (a fork) before the local (the prongs) so we don’t initially see a problem until we examine the object carefully to try and separate figure and ground
What are 3 models of object recognition
Template matching. Feature recognition. Structural theories
Discuss template matching
Compares the whole object to stored representations to find a match.
To compare the shape percept to memory for known shapes and making a decision about whether the object is familiar, and tap into knowledge about what can we use
3 Models of object recognition - template matching, feature recognition, structural theories
Discuss feature analysis
Rather than focusing on the whole shape, feature models break shapes down into critical features - those critical features are recognised and then assembled into objects and shapes that are compared to mental templates.
Discuss strengths of feature detection
Reduces the number of representations the mind must stroe in order to process everything to be recognised. Physiological evidence of feature detectors
Discuss the limitations of matching models in general
Based on 2D but we see 3D. It cannot account for superficial differences in the same class of item. For example, in handwriting, different types of duck, guitar or yak. Can’t account for learning of new instances - instead has to be presented with them.
Discuss Marr & Nishihara’s (1978) model
Proposed 3 different levels of representation underpinning object recognition: Edge image (primal sketch) - provides 2D description of main light-intensity changes, including information about edges contours and blobs, it is observer centered. 2.5D sketch - incorporates depth and orientation of surfaces, makes use of shading, texture, motion, binocular disparity and is also observer-centered. 3D model representation - three-dimensional object shape, relative positions and viewpoint independent/invariant
Discuss where recognition occurs
Single model axis - identify main axis of object. Component axis - then identify the axes of each of the smaller sub-portions. 3D model match - between the arrangement of components and a stored 3D model description to identify object
Discuss Biederman’s recognition by components theory (RBC)
Objects consist of combinations of geons - geometric icons, combinations of 36 basic shapes. Object recognition is viewpoint invariant - emphasises bottom-up processes. Region of concativity are particularly important in RBC. Most experiments performed with familiar objects that we are used to seeing from multiple angles - does this differ?
