Abstract: This paper addresses the problem of learning similarity-preserving binary codes for efficient similarity search in large-scale image collections. We formulate this problem in terms of finding a rotation of zero-centered data so as to minimize the quantization error of mapping this data to the vertices of a zero-centered binary hypercube, and propose a simple and efficient alternating minimization algorithm to accomplish this task. This algorithm, dubbed iterative quantization (ITQ), has connections to multi-class spectral clustering and to the orthogonal Procrustes problem, and it can be used both with unsupervised data embeddings such as PCA and supervised embeddings such as canonical correlation analysis (CCA). The resulting binary codes significantly outperform several other state-of-the-art methods. We also show that further performance improvements can result from transforming the data with a nonlinear kernel mapping prior to PCA or CCA. Finally, we demonstrate an application of ITQ to learning binary attributes or “classemes” on the ImageNet dataset.

Abstract: Brightness induction is the modulation of the perceived intensity of an area by the luminance of surrounding areas. From a purely computational perspective, we built a low-level computational model (CIWaM) of early sensory processing based on multi-resolution wavelets with the aim of replicating brightness and colour (Otazu et al., 2010, Journal of Vision, 10(12):5) induction effects. Furthermore, we successfully used the CIWaM architecture to define a computational saliency model (Murray et al, 2011, CVPR, 433-440; Vanrell et al, submitted to AVA/BMVA'12). From a biological perspective, neurophysiological evidence suggests that perceived brightness information may be explicitly represented in V1. In this work we investigate possible neural mechanisms that offer a plausible explanation for such effects. To this end, we consider the model by Z.Li (Li, 1999, Network:Comput. Neural Syst., 10, 187-212) which is based on biological data and focuses on the part of V1 responsible for contextual influences, namely, layer 2-3 pyramidal cells, interneurons, and horizontal intracortical connections. This model has proven to account for phenomena such as visual saliency, which share with brightness induction the relevant effect of contextual influences (the ones modelled by CIWaM). In the proposed model, the input to the network is derived from a complete multiscale and multiorientation wavelet decomposition taken from the computational model (CIWaM).
This model successfully accounts for well known pyschophysical effects (among them: the White's and modied White's effects, the Todorovic, Chevreul, achromatic ring patterns, and grating induction effects) for static contexts and also for brigthness induction in dynamic contexts defined by modulating the luminance of surrounding areas. From a methodological point of view, we conclude that the results obtained by the computational model (CIWaM) are compatible with the ones obtained by the neurodynamical model proposed here.

Abstract: Tone-mapping operators (TMO) are used to display high dynamic range(HDR) images in low dynamic range (LDR) displays. Many computational and biologically inspired approaches have been used in the literature, being many of them based on multiresolution decompositions. In this work, a simple two stage model for TMO is presented. The first stage is a novel multiresolution contrast decomposition, which is inspired in a pyramidal contrast decomposition (Peli, 1990 Journal of the Optical Society of America7(10), 2032-2040).
This novel multiresolution decomposition represents the Michelson contrast of the image at different spatial scales. This multiresolution contrast representation, applied on the intensity channel of an opponent colour decomposition, is processed by a non-linear saturating model of V1 neurons (Albrecht et al, 2002 Journal ofNeurophysiology 88(2) 888-913). This saturation model depends on the visual frequency, and it has been modified in order to include information from the extended Contrast Sensitivity Function (e-CSF) (Otazu et al, 2010 Journal ofVision10(12) 5).
A set of HDR images in Radiance RGBE format (from CIS HDR Photographic Survey and Greg Ward database) have been used to test the model, obtaining a set of LDR images. The resulting LDR images do not show the usual halo or color modification artifacts.

Abstract: We present a novel method for human action recognition (HAR) based on estimated poses from image sequences. We use 3D human pose data as additional information and propose a compact human pose representation, called a weak pose, in a low-dimensional space while still keeping the most discriminative information for a given pose. With predicted poses from image features, we map the problem from image feature space to pose space, where a Bag of Poses (BOP) model is learned for the final goal of HAR. The BOP model is a modified version of the classical bag of words pipeline by building the vocabulary based on the most representative weak poses for a given action. Compared with the standard k-means clustering, our vocabulary selection criteria is proven to be more efficient and robust against the inherent challenges of action recognition. Moreover, since for action recognition the ordering of the poses is discriminative, the BOP model incorporates temporal information: in essence, groups of consecutive poses are considered together when computing the vocabulary and assignment. We tested our method on two well-known datasets: HumanEva and IXMAS, to demonstrate that weak poses aid to improve action recognition accuracies. The proposed method is scene-independent and is comparable with the state-of-art method.

Abstract: State-of-the-art handwriting recognition systems are learning-based systems that require large sets of training data. The creation of training data, and consequently the creation of a well-performing recognition system, requires therefore a substantial amount of human work. This can be reduced with semi-supervised learning, which uses unlabeled text lines for training as well. Current approaches estimate the correct transcription of the unlabeled data via handwriting recognition which is not only extremely demanding as far as computational costs are concerned but also requires a good model of the target language. In this paper, we propose a different approach that makes use of keyword spotting, which is significantly faster and does not need any language model. In a set of experiments we demonstrate its superiority over existing approaches.

Abstract: In order to improve the results of automatically recognized handwritten text, information about the language is commonly included in the recognition process. A common approach is to represent a text line as a sequence. It is processed in one direction and the language information via n-grams is directly included in the decoding. This approach, however, only uses context on one side to estimate a word’s probability. Therefore, we propose a bidirectional recognition in this paper, using distinct forward and a backward language models. By combining decoding hypotheses from both directions, we achieve a significant increase in recognition accuracy for the off-line writer independent handwriting recognition task. Both language models are of the same type and can be estimated on the same corpus. Hence, the increase in recognition accuracy comes without any additional need for training data or language modeling complexity.

Abstract: In this paper, we propose a new approach to extract text regions from graphical documents. In our method, we first empirically construct two sequences of learned dictionaries for the text and graphical parts respectively. Then, we compute the sparse representations of all different sizes and non-overlapped document patches in these learned dictionaries. Based on these representations, each patch can be classified into the text or graphic category by comparing its reconstruction errors. Same-sized patches in one category are then merged together to define the corresponding text or graphic layers which are combined to createfinal text/graphic layer. Finally, in a post-processing step, text regions are further filtered out by using some learned thresholds.