Examinando por Autor "Cuesta Infante, Alfredo"

Mostrando 1 - 2 de 2

Bayesian capsule networks for 3D human pose estimation from single 2D images
(Neurocomputing, 2020) Ramírez Díaz, Iván; Cuesta Infante, Alfredo; Schiavi, Emanuele; Pantrigo Fernández, Juan José
Deep Bayesian Networks are a hot topic in Deep Learning because this approach makes it possible to minimize both the epistemic and the homoscedastic uncertainty at the same time self balancing multiple and complementary losses for a given task, simply by employing standard operations such as dropout, mean squared error or cross-entropy. On the other hand, Capsule networks are a novel DNN architecture that offer a richer representation because each concept is represented by a number of different vectors. The bayesian formulation of the Capsule networks is still an open problem that we address in this paper. We present a bayesian formulation of Capsule networks and compare its performance against the state-of-the-art for the ill-posed regression problem of estimating the 3D human pose from a single 2D image. The results show that our network is very competitive with a much more straightforward solution. To enable fair comparisons the source code is openly available at https://github.com/rollervan/BCN_3DPose/
Convolutional neural networks for computer vision-based detection and recognition of dumpsters
(Neural Computing and Applications, 2020) Ramírez Díaz, Iván; Cuesta Infante, Alfredo; Pantrigo Fernández, Juan José; Sanz Montemayor, Antonio; Moreno, José Luis; Alonso, Valvanera; Anguita, Gema; Palombarani, Luciano
In this paper, we propose a twofold methodology for visual detection and recognition of different types of city dumpsters, with minimal human labeling of the image data set. Firstly, we carry out transfer learning by using Google Inception-v3 convolutional neural network, which is retrained with only a small subset of labeled images out of the whole data set. This first classifier is then improved with a semi-supervised learning based on retraining for two more rounds, each one increasing the number of labeled images but without human supervision. We compare our approach against both to a baseline case, with no incremental retraining, and the best case, assuming we had a fully labeled data set. We use a data set of 27,624 labeled images of dumpsters provided by Ecoembes, a Spanish nonprofit organization that cares for the environment through recycling and the eco-design of packaging in Spain. Such a data set presents a number of challenges. As in other outdoor visual tasks, there are occluding objects such as vehicles, pedestrians and street furniture, as well as other dumpsters whenever they are placed in groups. In addition, dumpsters have different degrees of deterioration which may affect their shape and color. Finally, 35% of the images are classified according to the capacity of the container, which contains a feature which is hard to assess in a snapshot. Since the data set is fully labeled, we can compare our approach both against a baseline case, doing only the transfer learning using a minimal set of labeled images, and against the best case, using all the labels. The experiments show that the proposed system provides an accuracy of 88%, whereas in the best case it is 93%. In other words, the method proposed attains 94% of the best performance.