Flashcards
Object Representation
Flashcards do conjunto (21)
- 1Low Level Object RecognitionRetina sees world in contrasts not objects.
Low level signals - luminance, contrast, two types of colour contrast (Red / green and purple / yellow).
Spatial frequency channels crucial for obtaining information from different scenes at different scales, giving an idea of broad shape and fine detail - 2Spatial Filtering Through V1 FieldsDiffering sizes in excitatory centres and surrounds mean neurons are tuned to different features.
Neurons sensitive to different frequencies work in parallel, encoding the same image along different channels.
A.K.A. they all encode the same image, but detect different parts of it. - 3Spatial FrequencySpatial frequency - number of black and white stripes (cycles) occur per degree of visual angle.
As the viewing distance doubles, the size becomes smaller on the retina, so the spatial frequency doubles in cycles per degree (CPD). - 4Contrast Sensitivity Function CSFThe relation between contrast and spatial frequency, specifically how much contrast is required to perceive an object.
It measures an observer's sensitivity to contrast in stimuli of various sizes, making it a more comprehensive measure of visuo-spatial ability than visual acuity. - 5How Contrast Combines with Spatial FrequencySFC detect contrast, but are not all equally sensitive to it.
There are six spatial frequency channels (SFCs) representing the extent of spatial frequencies we can process. A-F.
These channels have bandwidths of 1 - 1.5 octaves.
We're optimally sensitive to mid-range frequencies (2-5 cpd) and can't resolve high SF patterns at a large distance (they look grey). - 6Why are Spatial Frequencies Important?Recognising faces (Brady & Oliva, 2012) and emotions (Kumar & Srinivasan, 2011)
To calculate the contrast of an image - it can be misleading. - 7How Do Spatial Frequencies Relate to Contrast?The Spatial Frequency Channel is sensitive to contrast within a narrow band of spatial frequencies.
- 8Spatial Frequencies in ObjectsFine detail and basic shape are produced in different pathways and at different speeds (Priebe, Lisberger & Movshon, 2006), with basic shape being processed more quickly.
Different SF scales produce different types of information about objects. - 9Spatial Frequencies in Objects (2)Not just cycles per degree on our retina, but cycles per object. Different distances used to detect different spatial frequencies of objects. Higher spatial frequencies become less visible at further distances.
- 10Reconciling Retinal and Object Based InformationDiagnostic information about an object that leads to its successful categorisation is contained in certain, critical SF bands.
The retinal projection of a stimulus changes with distance, and these changes regulate the SF information that will be available to vision.
Thus, category diagnostic, critical SF bands will be seen by different SF channels depending on viewing distance (Sowden & Schyns, 2006) - 11Bottom-Up and Top-Down InfluencesSmith, Cottrell, Gosselin and Schyns (2005) - we extract information about different emotions from different parts of the face, which can be more course or more global.
That means our bottom-up signal (contrast sensitivity) is modulated by our top-down signal (which point of the face to focus on).
Perceptual learning also occurs, where humans learn which SF channels carry diagnostic information for an object as a function of retinal image size. - 12ColourColour has less contrast sensitivity, but overlaps in four out of the six spatial frequency channels (A-D). We mainly use luminance to discern edges.
- 13Luminance and ColourLuminance is a better indicator of detail than colour, which does not aide object recognition (Wurm, Legge, Isenberg, & Luebker, 1993). When combined with luminance it can aide in scene recognition, however (Oliva & Schyns, 2000).
Largely, theories believe colour is useful for image segmentation and retrieval from memory, though this could be less to do with vision and more to do with emotional ties to colour. (Corridoni, Bimbo & Pala, 1999) - 14Mid/ High Level Object RepresentationMid - figures/ background, surfaces in 3D.
High - categorical knowledge. - 15Biederman's ModelObject recognition comes from components.
Problems - how do we distinguish similar objects? How can we recognise objects from some views not others (e.g. bottom of a salt shaker), little reference to rules of segmentation for large scenes.
This model is view-invarient - 16Ullman's Model - FragmentsView-dependent model
Highly informative, pictorial class fragments exist, where learned fragments relating to objects are stored and can be recalled and fitted together to aide object recognition.
These fragments are distinct and frequent. - 17Ullman's Model - HierarchiesThe fragments can be decomposed to get very fine-tuned hierarchies which enable us to generalise and diagnose what a new object is, due to us having created a highly nuanced representation. This helps us recognise objects by their individual parts, e.g. if an elf existed it would likely look human apart from the ears.
- 18Ullman's Model - Abstract Feature GroupingClass fragments allow us to deal with variability of objects within a class, allows for their grouping to form more abstract classes. Gives us a larger point of reference.
- 19Ullman's Model - Top-Down Class Based SegmentationTop-down processing - we focus our attention by starting with a larger concept and working our way to the finer details.
Images that are segmented using colour, motion and contrast can cause confusion due to a lack of specific criterion to begin with (how do you know where the legs end if you don't know what it it).
Informative class fragments tell you what you're segmenting for (e.g. legs), not applying specifics. That would be bottom-down processing, and would be too ambiguous without previous knowledge of the image. - 20Summary of Ullman's ModelHierarchy of abstract fragments are continuously extracted from examples
These fragments are class specific, non generic features, are learned based on usefulness and are pictorial, thus reflecting the appearance of object components. - 21Summary of Object RepresentationObject recognition is a highly adaptive process with iterative learning.
Low level filters provide information for object recognition at different scales.
Segmentation/ recognition/ classification - concurrent and interactive.
Informações
- Descrição
- FlashCards sobre Object Representation, criado por claire.macaulay em 06-12-2015.
- Criado por
- claire.macaulay
- Visualizações
- 0
- Tags
- Avaliação