Factorized Local Appearance Models
We propose a novel scheme for image-based object detection and localization by modeling the joint distribution of k-tuple salient point feature vectors which are factorized component wise after an independent component analysis (ICA). Furthermore, we use a distance-sensitive histograming technique for capturing spatial dependencies which enable us to model non-rigid objects as well as distortions caused by articulation.
Background & Objective: For appearance based object modeling in images, the choice of method is usually a trade-off determined by the nature of the application or the availability of computational resources. Existing object representation schemes provide models either for global features or for local features and their spatial relationships. With increased complexity, the latter provides higher modeling power and accuracy. Among various local appearance and structure models, there are those that assume rigidity of appearance and viewing angle, thus adopting more explicit models while others employ stochastic models and use probabilistic distance/matching metrics. Our objective is to model the high-order dependencies of local image structure by estimating the complete joint distribution of multiple salient point feature vectors using a density factorization approach.
Technical Discussion: We construct a probabilistic appearance model with an emphasis on the representation of non-rigid and approximate local image structures. We use joint histograms on k-tuples (k salient points) to enhance the modeling power for local dependency, while reducing the complexity by histogram factorization along the feature components. Although, the gain in modeling power of joint densities can increase the computational complexity, we propose histogram factorization based on independent component analysis to reduce the dimensionality dramatically, thus reducing the computation to a level that can be easily handled by today's personal computers. For modeling local structures, we use a distance-sensitive histograming technique. A clear advantage of the proposed method is the flexibility in modeling spatial relationships. Experiments have yielded promising results on robust object localization in cluttered scenes as well as image retrieval. Most recently we have adopted parametric models using mixture of Gaussians with resulting enhancements in performance.
| Technical Reports: | |
| Modeling High-Order Dependencies in Local Appearance Models | |
| Higher-Order Dependencies in Local Appearance Models | |
| Local Appearance-Based Models using High-Order Statistics of Image Features | |
| Joint Distribution of Local Image Features for Appearance Modeling | |
| Factorized Local Appearance Models | |
| Factorization for Probabilistic Local Appearance Models | |
| ICA-based Probabilistic Local Appearance Models | |
Technology Area: Computer Vision
Modification Date: September 12, 2007