Abstract: Bag-of-words model (BOW) is inspired by the text classification problem, where a document is represented by an unsorted set of contained words. Analogously, in the object categorization problem, an image is represented by an unsorted set of discrete visual words (BOVW). In these models, relations among visual words are performed after dictionary construction. However, close object regions can have far descriptions in the feature space, being grouped as different visual words. In this paper, we present a method for considering geometrical information of visual words in the dictionary construction step. Object interest regions are obtained by means of the Harris-Affine detector and then described using the SIFT descriptor. Afterward, a contextual-space and a feature-space are defined, and a merging process is used to fuse feature words based on their proximity in the contextual-space. Moreover, we use the Error Correcting Output Codes framework to learn the new dictionary in order to perform multi-class classification. Results show significant classification improvements when spatial information is taken into account in the dictionary construction step.

Abstract: Camera-based person detection is of paramount interest due to its potential applications. The task is diffcult because the great variety of backgrounds (scenarios, illumination) in which persons are present, as well as their intra-class variability (pose, clothe, occlusion). In fact, the class person is one of the included in the popular PASCAL visual object classes (VOC) challenge. A breakthrough for this challenge, regarding person detection, is due to Felzenszwalb et al. These authors proposed a part-based detector that relies on histograms of oriented gradients (HOG) and latent support vector machines (LatSVM) to learn a model of the whole human body and its constitutive parts, as well as their relative position. Since the approach of Felzenszwalb et al. appeared new variants have been proposed, usually giving rise to more complex models. In this paper, we focus on an issue that has not attracted suficient interest up to now. In particular, we refer to the fact that HOG is usually computed from RGB color space, but other possibilities exist and deserve the corresponding investigation. In this paper we challenge RGB space with the opponent color space (OPP), which is inspired in the human vision system.We will compute the HOG on top of OPP, then we train and test the part-based human classifer by Felzenszwalb et al. using PASCAL VOC challenge protocols and person database. Our experiments demonstrate that OPP outperforms RGB. We also investigate possible differences among types of scenarios: indoor, urban and countryside. Interestingly, our experiments suggest that the beneficts of OPP with respect to RGB mainly come for indoor and countryside scenarios, those in which the human visual system was designed by evolution.

Abstract: The main advantage of segmentation by snakes consists in its ability to incorporate smoothness constraints on the detected shapes that can occur. Likewise, we propose to model snakes with other properties that reflect the information provided about the object of interest in a different extent. We consider different kinds of snakes, those searching for contours with a certain direction, those preserving an object’s model, those seeking for symmetry, those expanding open, etc. The availability of such a collection of snakes allows not only the more complete use of the knowledge about the segmented object, but also to solve some problems of the existing snakes. Our experiments on segmentation of facial features justify the usefulness of snakes with different properties.

Abstract: The log-polar space variant representation, motivated by biological vision, has been widely studied in the literature. Its data reduction and invariance properties made it useful in many vision applications. However, due to its nature, it fails in preserving features in the periphery. In the current work, as an attempt to overcome this problem, we propose a novel space-variant representation. It is evaluated and proved to be better than the log-polar representation in preserving the peripheral information, crucial for on-board mobile vision applications. The evaluation is performed by comparing log-polar and the proposed representation once they are used for estimating dense optical flow.

Abstract: Existing state of the art optical flow approaches, which are evaluated on standard datasets such as Middlebury, not necessarily have a similar performance when evaluated on driving scenarios. This drop on performance is due to several challenges arising on real scenarios during driving. Towards this direction, in this paper, we propose a modification to the regularization term in a variational optical flow formulation, that notably improves the results, specially in driving scenarios. The proposed modification consists on using the Laplacian derivatives of flow components in the regularization term instead of gradients of flow components. We show the improvements in results on a standard real image sequences dataset (KITTI).

Abstract: Image classification is widely researched in the literature, where models based on Convolutional Neural Networks (CNNs) have provided better results. When data is not enough, CNN models tend to be overfitted. To deal with this, often, traditional techniques of data augmentation are applied, such as: affine transformations, adjusting the color balance, among others. However, we argue that some techniques of data augmentation may be more appropriate for some of the classes. In order to select the techniques that work best for particular class, we propose to explore the epistemic uncertainty for the samples within each class. From our experiments, we can observe that when the data augmentation is applied class-conditionally, we improve the results in terms of accuracy and also reduce the overall epistemic uncertainty. To summarize, in this paper we propose a class-conditional data augmentation procedure that allows us to obtain better results and improve robustness of the classification in the face of model uncertainty.

Abstract: Active learning aims to select samples to be annotated that yield the largest performance improvement for the learning algorithm. Many methods approach this problem by measuring the informativeness of samples and do this based on the certainty of the network predictions for samples. However, it is well-known that neural networks are overly confident about their prediction and are therefore an untrustworthy source to assess sample informativeness. In this paper, we propose a new informativeness-based active learning method. Our measure is derived from the learning dynamics of a neural network. More precisely we track the label assignment of the unlabeled data pool during the training of the algorithm. We capture the learning dynamics with a metric called label-dispersion, which is low when the network consistently assigns the same label to the sample during the training of the network and high when the assigned label changes frequently. We show that label-dispersion is a promising predictor of the uncertainty of the network, and show on two benchmark datasets that an active learning algorithm based on label-dispersion obtains excellent results.