Object recognition using point uncertainty regions as pose uncertainty regions
Section snippets
Introduction and related work
Model-based object recognition systems usually belong to one of three classes of algorithms. The first and most common type has been studied extensively and systems based on this concept have been presented for example in [1], [2], [3]. In this type of algorithm, which is known as the Alignment method, a minimal set of image features is matched to model features. From this hypothesized match the pose of the object is computed. The correctness of the match is verified by projecting the model on
Point uncertainty regions
In this section, we present a short overview on the topic of point uncertainty regions and how to estimate them. More details can be found in [18], [19].
Given a match of three model points (m0, m1, m2) to three points from an image under the weak-perspective projection model (i0, i1, i2), the pose of the object can be computed [17]. The pose is represented by two values H1 and H2, which give the tilt of the plane spanned by the three model points with respect to the image plane (see Fig. 2).
Proposed approach
One of the classical methods in artificial intelligence for solving problems is the ‘generate (hypothesis) and test’ method. In our setting, hypotheses are generated by matching triplets of model points to image points and are verified by projecting model points on the image and searching for points in the image within the uncertainty regions. Here a different approach is proposed. Hypotheses are generated in an order sorted by a priori probability (described in Section 4). The poses, which are
Probabilistic match ranking
A very important factor in the computational cost of running the algorithm is the order in which hypotheses are generated. A large computational speedup can be achieved when more probable hypotheses are tested first. To achieve this goal the ‘probabilistic peaking effect’ is exploited. This effect states that angles and ratios of lengths measured in the image will with high probability be close to the actual angles and ratios measured between their respective model points (Fig. 3). This means
Algorithm and implementation
We have implemented the algorithm and tested it on simulated and real images. In building such an implementation two issues have to be addressed in order to yield an efficient algorithm. First, the number of 3D points k has to be chosen and second, the structure of the lookup table has to be determined.
Experimental results
The algorithm was implemented and experiments were conducted to verify its applicability and test its various components.
In the first experiment the running of algorithm with and without the probabilistic ordering of hypotheses was compared. A thousand runs of the algorithm were performed for both cases. As can be seen from the runtime histograms shown in Fig. 5, the probabilistic ordering has a major impact on reducing the runtime of the algorithm. Under the random ordering of hypotheses, the
Conclusion
In this paper, we have presented a recognition algorithm, which searches for pairs of consistent matches between model and image points, which yield the same pose. Poses are stored in a table represented by the projections of several 3D points onto the image and their uncertainty regions, which represent the uncertainty in the pose. If all pairs of uncertainty regions intersect the two matches are declared consistent and yield very similar poses. This algorithm exploits the fact that we are
References (24)
Solution of the simultaneous pose and correspondence problem using gaussian error model
Comput. Vis. Image Underst.
(1999)Object recognition and localization via pose clustering
Comput. Vis. Graph. Image Proc.
(1987)- et al.
Object location by parallel pose clustering
Comput. Vis. Image Understand.
(1998) - et al.
Feature-based object recognition and localization in 3d-space, using a single video image
Comput. Vis. Image Understand.
(1999) - et al.
New algorithms for 2d and 3d point matching: pose estimation and correspondence
Pattern Recognit.
(1998) On estimating the uncertainty in the location of image points in 3D recognition from match sets of different sizes
Comput. Vis. Image Underst.
(1999)- et al.
Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography
Commun. ACM
(1981) - et al.
Recognizing 3D solid objects by alignment with an image
Int. J. Comput. Vis.
(1990) - V. Lepetit, J. Pilet, P. Fua, Point matching as a classification problem for fast and robust object pose estimation,...
- T. Cass, Feature matching for object localization in the presence of uncertainty, in: Proceedings of the International...
Probabilistic 3D object recognition
Int. J. Comput. Vis.
Cited by (1)
Application of adaptive Monte Carlo method to evaluate pose uncertainty in monocular vision system
2022, Optical Engineering
- 1
Ilan Shimshoni was supported in part by the Israeli Ministry of Science, Grant no. 2104. He was also supported by the fund for the promotion of research at the Technion.