Selecting Canonical Views for View-Based 3-D Object Recognition

In this paper we focus on automatically learning object models in the framework of keypoint based object recognition. The proposed method uses a collection of views of the objects to build the model. For each object the collection is composed of NxM views obtained rotating the object around its vertical and horizontal axis. As keypoint based object recognition using a complete set of views is computationally expensive, we focused on the definition of a selection method that creates, for each object, a subset of the initial views that visually summarize the characteristics of the object and should be suited for recognition. We select the views by determining maxima and minima of a function, based on the number of SIFT descriptors able to evaluate views similarity and relevance. Experimental results for recognition on a publicly available dataset are reported.

2004

Given a set of patterns and a similarity measure between them, we will present an optimization framework to approximate a small subset, known as a canonical set, whose members closely resemble the members of the original set. We will present a combinatorial formulation of the canonical set problem in terms of quadratic optimization integer programming, present a relaxation through semidefinite programming, and propose a bounded performance rounding procedure for its approximation solution in polynomial time. Through a set of experiments we will investigate the application of canonical sets for computing a summary of views from a dense set of 2D views computed for a 3D object.

2007 International Conference on Computing: Theory and Applications (ICCTA'07), 2007

In this work, we present a method for model-based recognition of 3d objects from a small number of 2d intensity images taken from nearby, but otherwise arbitrary viewpoints. Our method works by linearly combining images from two (or more) viewpoints of a 3d object to synthesise novel views of the object. The object is recognised in a target image by matching to such a synthesised, novel view. All that is required is the recovery of the linear combination parameters, and since we are working directly with pixel intensities, we suggest searching the parameter space using an evolutionary optimisation algorithm in order to efficiently recover the optimal parameters and thus recognise the object in the scene.

Pattern Recognition, 2000

We propose a new method for recognizing three-dimensional objects using a three-dimensional invariant relationship and geometric hashing by single-view. We develop a special structure consisting of four co-planar points and any two non-planar points with respect to the plane. We derive an invariant relationship for the structure, which is represented by a plane equation. For the recognition of three-dimensional objects using the geometric hashing, a set of points on the plane, thereby satisfying the invariant relationship, are mapped into a set of points intersecting the plane and the unit sphere. Since the structure is much more general than the previous structures proposed by Rothwell et al. (Oxford University gives enough many voting to generate hypotheses. We also show that from the proposed invariant relationship, an invariant for the structure by two-view and an invariant for a structure proposed by Zhu et al. (also be derived. Experiments using three-dimensional polyhedral objects are carried out to demonstrate the feasibility of our method for three-dimensional objects.

… Conference on Pattern …, 2010

In this paper we propose a new framework for view-invariant 3D object recognition, based on what we call Visibility Maps. A Visibility Map (VM) encodes a compact model of an arbitrary 3D object for which a set of images taken from different views is available. Representative local invariant features extracted from each image are selectively combined to form a visibility basis, in terms of which an arbitrary view of the modeled object can be represented. A metric which incorporates geometric information is also provided for comparing test images to the model, and can be used for recognition.

Image and Vision Computing, 1992

2005

This article introduces a novel representation for three-dimensional (3D) objects in terms of local affine-invariant descriptors of their images and the spatial relationships between the corresponding surface patches. Geometric constraints associated with different views of the same patches under affine projection are combined with a normalized representation of their appearance to guide matching and reconstruction, allowing the acquisition of true 3D affine and Euclidean models from multiple unregistered images, as well as their recognition in photographs taken from arbitrary viewpoints. The proposed approach does not require a separate segmentation stage, and it is applicable to highly cluttered scenes. Modeling and recognition results are presented.

IEEE Journal on Robotics and Automation, 1987

A robotic vision system is being developed which uses threedimensional laser range data to sense its environment. The recognition subsystem incorporates topological as well as geometric information to identify viewed objects. Theorem-proving techniques are used to produce symbolic pattern matches. The major contributions of the recognition subsystem are 1) the use of viewpoint independent descriptors as the basis for representing known object models and 2) the use of theorem proving techniques to hypothesize object identities and recognize the viewed object as an instance of the appropriate viewpoint independent model descriptor. The representation scheme permits describing objects at a variety of topological and geometric levels. Furthermore, the use of viewpoint independent descriptors facilitates object recognition from a single arbitrary view despite missing information or the inclusion of viewpoint dependent artifacts. The theorem-proving approach establishes a symbolic correspondence between viewpoint independent features in the (recognized) model and features in the observed data. The recognition process uses a three-phase approach. First, hypotheses are generated which correspond to model descriptors that are likely to match the data. Evidence is applied to viable hypotheses to produce a partial match. The partial match is then used to constrain the full recognition process which leads to object identification. This strategy has been found to constrain strongly the search space of possible matches and leads to large reductions in recognition times. Results of the recognition process on synthetic and actual laser range data are presented for several objects. The system is shown to operate with robustness and alacrity.

This thesis addresses the problem of learning object classification using multi-view range data. Class membership is determined by shared characteristics, which can be visual, structural or functional. The major steps in object recognition and object classification are: segmentation, feature extraction, object representation and learning.

dpl.ceegs.ohio-state.edu

This paper introduces a method for the recognition planar objects under projective geometry. Our method is based on a similarity measure invariant to projective transform. The proposed similarity measure utilizes the distribution of the projective relations between the conic section pairs, which are estimated from the object's shape. We conjecture that given two objects of the same type, which are viewed from different viewpoints generate similar histograms, such that their difference is smaller than the histograms generated from other object types. The proposed measure has shown promising performance on the Brown shape database.