This work analyzes discriminative capabilities of a hierarchical compositional model. A poor discriminative performance has been shown to stem from excessive feature sharing and generative learning that does not prioritize discriminative features. Histogram of Compositions (HoC) is introduced as a viable solution to address this issue. HoC improves the discriminative power by merging the high-level detection information with the low-level features that are important for discrimination. We apply our solution to the problem of object detection using shape-specific features, where an extensive evaluation on five datasets shows a significant improvement of the overall detection performance. Improved compositional hierarchical model is also compared against a state-of-the-art deep convolutional network, R-CNN. Under ideal conditions R-CNN is shown to perform better, but is on the other hand much more sensitive to various types of heavy occlusions compared to compositional hierarchical model with improved discriminative power.
COBISS.SI-ID: 1536363971
In this paper, we present a comprehensive performance evaluation of several batch and on-line multi-class formulations of linear Support Vector Machine classifiers, applied to the domain of traffic sign recognition. Using different types of features, we show that the on-line/incremental formulations of the classifiers offer performance that is consistently on par performance of their batch counterparts. Furthermore, we demonstrate the benefit of incremental learning in an on-line learning scenario, where the classifier continues to be updated with new samples. The obtained results indicate that the on-line classifiers are a suitable replacement for batch methods, as well as a natural choice for open-world recognition problems that require constant adaptation of the system.
COBISS.SI-ID: 1536821443
This work proposes an extension to deep convolutional networks to explicitly model spatial structures that are commonly present in visual objects. A restriction between a feature and its sub-features is introduced to explicitly represent the spatial structure within the object’s hierarchy. Spatial restriction is imposed through parametrization of filters using a mixture of Gaussian distributions. We show how this model fits into a standard learning procedure of deep neural networks using back-propagation and gradient descent learning. We show that a standard convolutional neural network can be considered as a special case of the proposed model on one hand, while more importantly on the other hand, proposed model is show to be now considered as a compositional hierarchy as well. An evaluation of the proposed model is performed on a classification problem with two datasets, which shows the proposed model achieving a comparable performance to the deep networks, while attaining a higher level of compactness of filters, thus making a successful first step towards the compositional interpretation of the representations being learned by CNNs.