In our paper we have proposed a novel approach to analysis of hierarchical compositional models that preserves the appealing properties of the generative hierarchical models, while at the same time improves their discrimination properties. We achieve this by introducing a network of discriminative nodes on top of the existing generative hierarchy. The discriminative nodes are sparse linear combinations of activated generative parts. We show in the experiments that the discriminative nodes consistently improve a state-of-the-art hierarchical compositional model. Results show that our approach considers only a fraction of all nodes in the vocabulary (less than 10%) which also makes the system computationally efficient.
COBISS.SI-ID: 9952852
We proposed a novel approach to online estimation of generative models, which is based on probability density estimation. As the theoretical framework we use kernel density estimation (KDE). The method maintains and updates a nonparametric model of observed data, from which the KDE can be calculated. We propose an online bandwidth estimation approach and a compression/revitalization scheme. The results show that the online KDE outperforms or achieves a comparable performance to the stateoftheart and produces models with a significantly lower complexity while allowing online adaptation.
COBISS.SI-ID: 8289876
In order for recognition systems to scale to a larger number of object categories building visual class taxonomies is important to achieve running times logarithmic in the number of classes. In this paper we propose a novel approach for speeding up recognition times of multiclass partbased object representations. The main idea is to construct a taxonomy of constellation models cascaded from coarsetofine resolution and use it in recognition with an efficient search strategy. The approach is utilized on the hierarchy-of-parts model.
COBISS.SI-ID: 8255828
We proposed an architecture for an object detection system suitable for a web-service running distributed on a cluster of machines. By implementing hierarchical compositional models in distributed system we enabled scalability of the algorithm to higher number of images required to process higher number of categories with more images. We implemented distributed system by utilizing multi-core processing units (multi-core CPUs) and by utilizing platforms for big data processing spanning over multiple machines (cluster). As part of distributed big data processing we deployed learning algorithms of hierarchical compositional models on a Hadoop platform and algorithms for object detection on a Storm platform. Storm implementation for object detection is also part of an publicly available web-service ViCoS-Eye.
COBISS.SI-ID: 10008916
We have proposed a new approach to tracking using a hierarchical visual model. We developed a new probabilistic model for tracking nonrigid objects. The model is composed of two layers. The top layer consists of a global information about the appearance of the object, while the lower layer contains the object’s local information. The local layer consists of a vocabulary of basic visual elements. These visual elements are weakly coupled in a constellation that can deform during the tracking. For this purpose, we have developed a new statistical model and have proposed an optimization method, which efficiently adapts to the target’s appearance through significant deformation and partial occlusion. On the one hand, the proposed model is a test-bed for application of parts of theory of the hierarchical models to the case of object tracking. On the other hand, we developed a robust tracking algorithm that will allow us to capture large datasets of articulated objects. These will be used as an input to the object category detection hierarchy that we are developing. The journal in which this work was published is considered the top journal in computer vision.
COBISS.SI-ID: 9431124