The Cambridge Handbook of Artificial Intelligence by Unknown

Author:Unknown
Language: eng
Format: epub
Publisher: Cambridge University Press
Published: 2014-06-30T00:00:00+00:00

8.2.1 2D modeling

Most objects in the real world are inherently 3D. Nevertheless, many object-recognition techniques stick to 2D representations with significant success. There are several reasons for this. (1) Easy accessibility: We get 2D image information nearly for free using standard camera equipment. (2) Fast computation: Features can directly be calculated from image pixel data and do not involve a search for complex geometric primitives. (3) Simple acquisition of detection models: Models that are used for automatic object detection are typically learned from example images. (4) Robustness to noise: Features are directly computed on pixel values. This is in contrast to the extraction of more abstract primitives (regions, contours, 3D shape primitives) that typically involves segmentation issues and, therefore, is more error prone with regard to clutter and noise. (5) Furthermore, many interesting objects have quite characteristic 2D views – for example, cover pages, traffic signs, side views of motor bikes or cars, front views of faces.

The price to pay for ignoring the 3D characteristics of objects is typically over- or under-constrained models because there are a number of perspective variations that cannot be systematically dealt with. A typical case of under-constrained approaches are bag-of-feature models. Like the histogram models mentioned in Section 8.1, these compute feature statistics over an image region or complete image. Thus, the location of features is completely lost, and object rotation and exact position cannot be distinguished. Thus, for example, if the eyes, nose, and mouth of a face were upside down or completely intermixed, the recognizer would still wrongly detect a face. Over-constrained models, on the other hand, need multiple representations in order to deal with different part configurations or rotations of objects. (Good examples are the template-based methods mentioned before.) Hence if, for example, a face is rotated by 90 degrees, the recognizer would never detect it. As an additional price to pay, we need to cope with a more challenging segmentation problem – that is, the problem of extracting an object from its background. Typically, the background is further away, so that 3D information provides a much stronger hint than luminance values of 2D images.

The dominant class of 2D object recognition techniques are appearance-based approaches. Instead of using a view-invariant object-centered representation, these represent different aspects of an object. Compact representations are provided by aspect-graphs (Koenderink 1987) that relate different 2D appearances to each other in an efficient data structure. Secondly, appearance-based approaches drop an intermediate geometric representation level by computing features directly from pixel values. This has certain consequences for the kind of object classes that can be distinguished and the within-class variations that can be covered.

So far, the methods discussed deal with variations of rotation, lighting, noise, and small distortions of an object’s shape. They mostly assume that objects are solid, approximately rigid, have similar textures or colors, and are occluded to a minor degree. Further variations are covered by local descriptor approaches. Here, the main idea is to detect salient points in an image that provide a partial feature description instead of a complete appearance model.

Download

Copyright Disclaimer:
This site does not store any files on its server. We only index and link to content provided by other sites. Please contact the content providers to delete copyright contents if any and email us, we'll remove relevant links or contents immediately.