Recognition by Linear Combinations of Models

By Shimon Ullman, Ronen Basri. 1991

An approach to visual object recognition in which a 3D object is represented by the linear combination of 2D images of the object is proposed. It is shown that for objects with sharp edges as well as with smooth bounding contours, the set of possible images of a given object is embedded in a linear space spanned by a small number of views. For objects with sharp edges, the linear combination representation is exact. For objects with smooth boundaries, it is an approximation that often holds over a wide range of viewing angles. Rigid transformations (with or without scaling) can be distinguished from more general linear transformations of the object by testing certain constraints placed on the coefficients of the linear combinations. Three alternative methods of determining the transformation that matches a model to a given image are proposed. >

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

Citations in the corpus

Fast Edge Detection Using Structured Forests (2014, neutral): “F”
Unsupervised Learning of Edges (2015, background): “Endowing machine vision systems with similar capabilities is of interest as edges are useful for diverse tasks such as optical flow [31], object detection [40, 13], and object proposals [39, 46, 3].”
Non-negative tensor factorization with applications to statistics and computer vision (2005, background): “( Ullman & Basri, 1991 ) — therefore, a minimum of 5 matching points are necessary for verifying whether they come from the same body.”
A multi-view probabilistic model for 3D object classes (2009, methodology): “Notably, [3, 32 , 8, 24] have proposed methods for identifying and recognizing single objects such as a cereal box or a Teddy bear, under arbitrary viewing conditions.”
Visual learning and recognition of 3-d objects from appearance (2005, background): “UUman and Basri ( Ullman and Basri 1991 ) showed that three views of an object can be used to represent its boundaries.”
Single image 3D object detection and pose estimation for grasping (2014, background): “Ullman [17] represented a 3D object by a linear combination of a small number of images enabling an alignment of the unknown object with a model by computing the coefficients of the linear combination, and, thus, reducing the problem to 2D.”
Multidimensional morphable models (1998, result): “Several of them represent objects using 3D models, represented in di erent ways (for a review, see [1]).”
Sketch Tokens: A Learned Mid-level Representation for Contour and Object Detection (2013, methodology): “Early edge detectors [3] were used to ﬁnd more complex shapes such as junctions [22], straight lines and curves [8], and were applied to object recognition [34], structure from motion [32], tracking [31], and 3D shape recovery [25].”
In search of illumination invariants (2001, background): “For example, Jacobs, Belhumeur, and Basri [24] show that under arbitrary, di use lighting conditions, it is not possible to exactly reconstruct the albedos of an object, even when its structure is precisely known.”
Structured Forests for Fast Edge Detection (2013, background): “The detection of edges is a critical preprocessing step for a variety of tasks, including object recognition [36, 12], segmentation [23, 1], and active contours [19].”
Lambertian reflectance and linear subspaces (2001, background): “[25] analytically derives such a representation for sets of 3D points undergoing scaled orthographic projection.”
Looking at People: Sensing for Ubiquitous and Wearable Computing (2000, background): “…colors lie on a one-dimensional manifold (with color variation primarily due to melanin concentration), that the unusual spatio-temporal structure of eye blinks that makes them easy to detect, or that human facial geometry is limited and essentially 2D when people are looking toward the camera.”
Part-based modelling of compound scenes from images (2015, methodology): “This approach is thus more closely related to work such as [21] which describes a method for object recognition using linear combinations of images.”
Face recognition: A literature survey (2003, background): “person are available [191, 192, 193, 194], 2) hybrid methods when multiple images are available during training but only one database image per person is available during recognition [80, 195, 196, 197], and 3) single image based methods when no training is carried out.”
Task-Specific Gesture Analysis in Real-Time Using Interpolated Views (1996, methodology): “This approach is related to the idea of view-based representation, as advocated by Ullman [18] and Poggio [14], for representing 3-D objects by interpolating between a small set of 2-D views.”
Learning viewpoint invariant object representations using a temporal coherence principle (2005, background): ““Image-based” or “view-based” models in turn postulate the existence of a small number of two-dimensional templates for objects and propose several mechanisms that could allow generalization to novel viewpoints, such as mental rotation (Tarr and Pinker 1989), mapping to a standard (“canonical”) view”