Abstract: The simultaneous use of images from different spectracan be helpful to improve the performance of many computer vision tasks. The core idea behind the usage of crossspectral approaches is to take advantage of the strengths of each spectral band providing a richer representation of a scene, which cannot be obtained with just images from one spectral band. In this work we tackle the cross-spectral image similarity problem by using Convolutional Neural Networks (CNNs). We explore three different CNN architectures to compare the similarity of cross-spectral image patches. Specifically, we train each network with images from the visible and the near-infrared spectrum, and then test the result with two public cross-spectral datasets. Experimental results show that CNN approaches outperform the current state-of-art on both cross-spectral datasets. Additionally, our experiments show that some CNN architectures are capable of generalizing between different crossspectral domains.

Abstract: In order to deal with the huge amount of content generated by social media, especially for indexing and retrieval purposes, the focus shifted from single object recognition to multi-observation object recognition. Of particular interest is the problem of face recognition (used as primary cue for persons’ identity assessment), since it is highly required by popular social media search engines like Facebook and Youtube. Recently, several approaches for graph-based label propagation were proposed. However, the associated graphs were constructed in an ad-hoc manner (e.g., using the KNN graph) that cannot cope properly with the rapid and frequent changes in data appearance, a phenomenon intrinsically related with video sequences. In this paper, we
propose a novel approach for efficient and adaptive graph construction, based on a two-phase scheme: (i) the first phase is used to adaptively find the neighbors of a sample and also to find the adequate weights for the minimization function of the second phase; (ii) in the second phase, the
selected neighbors along with their corresponding weights are used to locally and collaboratively estimate the sparse affinity matrix weights. Experimental results performed on Honda Video Database (HVDB) and a subset of video
sequences extracted from the popular TV-series ’Friends’ show a distinct advantage of the proposed method over the existing standard graph construction methods.

Abstract: In this paper, we propose a novel way to perform top-down semantic feedback in convolutional deep networks for efficient and accurate image parsing. We also show how to add global appearance/semantic features, which have shown to improve image parsing performance in state-of-the-art methods, and was not present in previous convolutional approaches. The proposed method is characterised by an efficient training and a sufficiently fast testing. We use the well known SIFTflow dataset to numerically show the advantages provided by our contributions, and to compare with state-of-the-art image parsing convolutional based approaches.

Abstract: Manifold learning techniques are affected by two critical aspects: (i) the design of the adjacency graphs, and (ii) the embedding of new test data---the out-of-sample problem. For the first aspect, the proposed schemes were heuristically driven. For the second aspect, the difficulty resides in finding an accurate mapping that transfers unseen data samples into an existing manifold. Past works addressing these two aspects were heavily parametric in the sense that the optimal performance is only reached for a suitable parameter choice that should be known in advance. In this paper, we demonstrate that sparse coding theory not only serves for automatic graph reconstruction as shown in recent works, but also represents an accurate alternative for out-of-sample embedding. Considering for a case study the Laplacian Eigenmaps, we applied our method to the face recognition problem. To evaluate the effectiveness of the proposed out-of-sample embedding, experiments are conducted using the k-nearest neighbor (KNN) and Kernel Support Vector Machines (KSVM) classifiers on four public face databases. The experimental results show that the proposed model is able to achieve high categorization effectiveness as well as high consistency with non-linear embeddings/manifolds obtained in batch modes.

Abstract: Training vision-based pedestrian detectors using synthetic datasets (virtual world) is a useful technique to collect automatically the training examples with their pixel-wise ground truth. However, as it is often the case, these detectors must operate in real-world images, experiencing a significant drop of their performance. In fact, this effect also occurs among different real-world datasets, i.e. detectors' accuracy drops when the training data (source domain) and the application scenario (target domain) have inherent differences. Therefore, in order to avoid this problem, it is required to adapt the detector trained with synthetic data to operate in the real-world scenario. In this paper, we propose a domain adaptation approach based on boosting LDA exemplar classifiers from both virtual and real worlds. We evaluate our proposal on multiple real-world pedestrian detection datasets. The results show that our method can efficiently adapt the exemplar classifiers from virtual to real world, avoiding drops in average precision over the 15%.

Abstract: This paper presents an automatic 3D head pose initialization scheme for a real-time face tracker with application to human-robot interaction. It has two main contributions. First, we propose an automatic 3D head pose and person specific face shape estimation, based on a 3D deformable model. The proposed approach serves to initialize our realtime 3D face tracker. What makes this contribution very attractive is that the initialization step can cope with faces
under arbitrary pose, so it is not limited only to near-frontal views. Second, the previous framework is used to develop an application in which the orientation of an AIBO’s camera can be controlled through the imitation of user’s head pose.
In our scenario, this application is used to build panoramic images from overlapping snapshots. Experiments on real videos confirm the robustness and usefulness of the proposed methods.

Abstract: Laughter detection is an important area of interest in the Affective Computing and Human-computer Interaction fields. In this paper, we propose a multi-modal methodology based on the fusion of audio and visual cues to deal with the laughter recognition problem in face-to-face conversations. The audio features are extracted from the spectogram and the video features are obtained estimating the mouth movement degree and using a smile and laughter classifier. Finally, the multi-modal cues are included in a sequential classifier. Results over videos from the public discussion blog of the New York Times show that both types of features perform better when considered together by the classifier. Moreover, the sequential methodology shows to significantly outperform the results obtained by an Adaboost classifier.