Search

Key-frame extraction for first-person vision (FPV) videos is a core technology for selecting important scenes and memorizing impressive life experiences in our daily activities. The difficulty of selecting key frames is the scene instability caused by head-mounted cameras used for capturing FPV videos. Because head-mounted cameras tend to frequently shake, the frames in an FPV video are noisier than those in a third-person vision (TPV) video. However, most existing algorithms for key-frame extraction mainly focus on handling the stable scenes in TPV videos. The technical development of key-frame extraction techniques for noisy FPV videos is currently immature. Moreover, most key-frame extraction algorithms mainly use visual information from FPV videos, even though our visual experience in daily activities is associated with human motions. To incorporate the features of dynamically changing scenes in FPV videos into our methods, integrating motions with visual scenes is essential. In this paper, we propose a novel key-frame extraction method for FPV videos that uses multi-modal sensor signals to reduce noise and detect salient activities via projecting multi-modal sensor signals onto a common space by canonical correlation analysis (CCA). We show that the two proposed multi-sensor integration models for key-frame extraction (a sparse-based model and a graph-based model) work well on the common space. The experimental results obtained using various datasets suggest that the proposed key-frame extraction techniques improve the precision of extraction and the coverage of entire video sequences.

Image interpolation is often implemented using one of two methods: optical flow or convolutional neural networks. These methods are typically pixel-based; they do not work well on objects between images far apart. Because they either rely on a simple frame average or pixel motion, they do not have the required knowledge of the semantic structure of the data. In this paper, we propose a method for image interpolation based on latent representations. We use a simple network structure based on a variational autoencoder and an adjustable hyperparameter that imposes the latent space distribution to generate accurate interpolation. To visualize the effects of the proposed approach, we evaluate a synthetic dataset. We demonstrate that our method outperforms both pixel-based methods and a conventional variational autoencoder, with particular improvements in nonsuccessive images.

Saliency mapping is an efficient means of processing large amounts of incoming visual information from images and videos. Existing methods based on feature integration and construction error provide a rough saliency mapping and are sensitive to real-world noise. Furthermore, for first-person-vision video related to human activities, existing methods identify objects in the image without considering the human actions being performed. To address this issue, we propose a novel supervised saliency mapping based on a sparse coding framework. Different from the normal sparse representation, which uses dictionary atoms to represent signals, we use the inverse expression, whereby the use original image or video frame signals provide the normal dictionary, and these signals are used to represent the salience superpixel matrix, which is learned from a class of training images. We then construct the salient map by inverse sparse coding. In this paper, we first describe how the salience superpixel matrix is extracted by supervised selection, and then explain how to obtain the salient map through an inverse sparse coding framework. The experimental results show the enhanced accuracy of the proposed method compared with previous methods, as well as the time efficiency of the proposed method.

Electroencephalogram (EEG)-based brain-computer interfaces (BCIs) can potentially enable people to non-invasively and directly communicate with others using brain activities. Artifacts generated from body activities (e.g., eyeblinks and teeth clenches) often contaminate EEGs and make EEG-based classification/identification hard. Although independent component analysis (ICA) is the gold-standard technique for attenuating the effects of such contamination, the estimated independent components are still mixed with artifactual and neuronal information because ICA relies only on the independence assumption. The same problem occurs when using independent vector analysis (IVA), an extended ICA method. To solve this problem, we designed an independent low-rank matrix analysis (ILRMA)-based automatic artifact reduction technique that clearly models sources from observations under the independence assumption and a low-rank nature in the frequency domain. For automatic artifact reduction, we combined the signal separation technique with an independent component classifier for EEGs named ICLabel. To assess the comparative efficiency of the proposed method, the discriminabilities of artifact-reduced EEGs using ICA, IVA, and ILRMA were determined using an open-access EEG dataset named OpenBMI, which contains EEG data obtained through three BCI paradigms [motor-imagery (MI), event-related potential (ERP), and steady-state visual evoked potential (SSVEP)]. BCI performances were obtained using these three paradigms after applying artifact reduction techniques, and the results suggested that our proposed method has the potential to achieve higher discriminability than ICA and IVA for BCIs. In addition, artifact reduction using the ILRMA approach clearly improved (by over 70%) the averaged BCI performances using artifact-reduced data sufficiently for most needs of the BCI community. The extension of ICA families to supervised separation that leaves the discriminative ability would further improve the usability of BCIs for real-life environments in which artifacts frequently contaminate EEGs.

Facing the increasing quantity of AI models applications, especially in life- and property-related fields, it is crucial for designers to construct safety- and security-critical systems. As a major factor affecting the safety of AI models, corner case data and its related description/detection techniques are important in the AI design phase and quality assurance. In this paper, inspired by surprise adequacy (SA), a tool having advantages on capture data behaviors, we developed three modified versions of distance-based-SA (DSA) for detecting corner cases in classification problems. Through the experiment analysis on MNIST, CIFAR, and industrial example data, the feasibility and usefulness of the proposed tools on corner case data detection are verified. Moreover, Qualitative and quantitative experiments validated that the developed DSA tools can achieve improved performance in describing corner cases’ behaviors.

Humans divide perceived continuous information into segments to facilitate recognition. For example, humans can segment speech waves into recognizable morphemes. Analogously, continuous motions are segmented into recognizable unit actions. People can divide continuous information into segments without using explicit segment points. This capacity for unsupervised segmentation is also useful for robots, because it enables them to flexibly learn languages, gestures, and actions. In this paper, we propose a Gaussian process-hidden semi-Markov model (GP-HSMM) that can divide continuous time series data into segments in an unsupervised manner. Our proposed method consists of a generative model based on the hidden semi-Markov model (HSMM), the emission distributions of which are Gaussian processes (GPs). Continuous time series data is generated by connecting segments generated by the GP. Segmentation can be achieved by using forward filtering-backward sampling to estimate the model's parameters, including the lengths and classes of the segments. In an experiment using the CMU motion capture dataset, we tested GP-HSMM with motion capture data containing simple exercise motions; the results of this experiment showed that the proposed GP-HSMM was comparable with other methods. We also conducted an experiment using karate motion capture data, which is more complex than exercise motion capture data; in this experiment, the segmentation accuracy of GP-HSMM was 0.92, which outperformed other methods.

There is growing interest in understanding how the brain utilizes synchronized oscillatory activity to integrate information across functionally connected regions. Computing phase-locking values (PLV) between EEG signals is a popular method for quantifying such synchronizations and elucidating their role in cognitive tasks. However, high-dimensionality in PLV data incurs a serious multiple testing problem. Standard multiple testing methods in neuroimaging research (e.g., false discovery rate, FDR) suffer severe loss of power, because they fail to exploit complex dependence structure between hypotheses that vary in spectral, temporal and spatial dimension. Previously, we showed that a hierarchical FDR and optimal discovery procedures could be effectively applied for PLV analysis to provide better power than FDR. In this article, we revisit the multiple comparison problem from a new Empirical Bayes perspective and propose the application of the local FDR method (locFDR; Efron, 2001) for PLV synchrony analysis to compute FDR as a posterior probability that an observed statistic belongs to a null hypothesis. We demonstrate the application of Efron's Empirical Bayes approach for PLV synchrony analysis for the first time. We use simulations to validate the specificity and sensitivity of locFDR and a real EEG dataset from a visual search study for experimental validation. We also compare locFDR with hierarchical FDR and optimal discovery procedures in both simulation and experimental analyses. Our simulation results showed that the locFDR can effectively control false positives without compromising on the power of PLV synchrony inference. Our results from the application locFDR on experiment data detected more significant discoveries than our previously proposed methods whereas the standard FDR method failed to detect any significant discoveries.

A method of detecting speech events in a multiple-sound-source condition using audio and video information is proposed. For detecting speech events, sound localization using a microphone array and human tracking by stereo vision is combined by a Bayesian network. From the inference results of the Bayesian network, information on the time and location of speech events can be known. The information on the detected speech events is then utilized in the robust speech interface. A maximum likelihood adaptive beamformer is employed as a preprocessor of the speech recognizer to separate the speech signal from environmental noise. The coefficients of the beamformer are kept updated based on the information of the speech events. The information on the speech events is also used by the speech recognizer for extracting the speech segment.

The problem of tracking multiple intermittently speaking speakers is difficult as some distinct problems must be addressed. The number of active speakers must be estimated, these active speakers must be identified, and the locations of all speakers including inactive speakers must be tracked. In this paper we propose a method for tracking intermittently speaking multiple speakers using a particle filter. In the proposed algorithm the number of active speakers is firstly estimated based on the Exponential Fitting Test (EFT), a source number estimation technique which we have proposed. The locations of the speakers are then tracked using a particle filtering framework within which the decomposed likelihood is used in order to decouple the observed audio signal and associate each element of the decomposed signal with an active speaker. The tracking accuracy is then further improved by the inclusion of a silence region detection step and estimation of the noise-only covariance matrix. The method was evaluated using live recordings of 3 speakers and the results show that the method produces highly accurate tracking results.

Multivariate analysis techniques are widely used to analyze high-dimensional data in many fields of scientific research. In most cases, however, analysis is conducted under the assumption that there is a priori knowledge of the form of the latent structures in the data. Recently, Kemp and Tenenbaum proposed a new method of analysis that can select the forms from a set of several primitive forms and structures from a set of latent structures for those forms. It is important to evaluate the validity and the effectiveness of the proposed method by using synthetic data sets so that we can control their form. In this study, we apply the Kemp-Tenenbaum method to synthetic data sets that are artificially generated by an associative memory model. The forms and the structures that had been embedded in the data sets were successfully reconstructed, which demonstrates the validity of the Kemp-Tenenbaum method. [ABSTRACT FROM AUTHOR]

A method of detecting speech events in a multiple-sound-source condition using audio and video information is proposed. For detecting speech events, sound localization using a microphone array and human tracking by stereo vision is combined by a Bayesian network. From the inference results of the Bayesian network, information on the time and location of speech events can be known. The information on the detected speech events is then utilized in the robust speech interface. A maximum likelihood adaptive beamformer is employed as a preprocessor of the speech recognizer to separate the speech signal from environmental noise. The coefficients of the beamformer are kept updated based on the information of the speech events. The information on the speech events is also used by the speech recognizer for extracting the speech segment. [ABSTRACT FROM AUTHOR]