Audio-Visual Event Localization by Learning Spatial and Semantic Co-Attention

With an overflow of multimedia information around us and an urgent need to identify data accurately, an audio and visual identification system with a high accuracy rate is developed to meet the demand. Classification and feature extraction are performed separately on audio and visual signals. Pending on the temporal correlation of the feature vectors of objects and speakers, indexes of all objects included in an audio/visual sequence are listed in a time sequence. In integrating the audio/visual features, every object or character of the key frames has a set of feature vectors; the user can select and search specific characters that have the audio and visual features from the entire index set. Due to integrating the audio/visual identification results in the time order, the proposed identification system can increase the accuracy about 4% and 6% in our experiments, comparing with the results using the audio features and visual features separately, respectively.

Automatic speech recognition by machine is an attractive research topic in signal processing domain and has attracted many researchers to contribute in this area. In recent year, there have been many advances in automatic speech reading system with the inclusion of audio and visual speech features to recognize words under noisy conditions. The objective of audio-visual speech recognition system is to improve recognition accuracy. In this paper we computed visual features using Zernike moments and audio feature using mel frequency cepstral coefficients on visual vocabulary of independent standard words dataset which contains collection of isolated set of city names of ten speakers. The visual features were normalized and dimension of features set was reduced by principal component analysis (PCA) in order to recognize the isolated word utterance on PCA space. The performance of recognition of isolated words based on visual only and audio only features results in 63.88 and 100 % respectively.

Human perception systems can integrate audio and visual information automatically to obtain a profound understanding of real-world events. Accordingly, fusing audio and visual contents is important to solve the audio-visual event (AVE) localization problem. Although most existing works have fused audio and visual modalities to explore their relationship with attention-based networks, we can delve into their relationship more deeply to improve the fusion capability of the two modalities. In this paper, we propose a dense modality interaction network (DMIN) to elegantly leverage audio and visual information by integrating two novel modules, namely, the audio-guided triplet attention (AGTA) module and the dense inter-modality attention (DIMA) module. The AGTA module enables audio information to guide the network to pay more attention to event-relevant visual regions. This guidance is conducted in the channel, temporal, and spatial dimensions, which emphasize informative features, temporal relationships and spatial regions, to boost the capacity of representations. Furthermore, the DIMA module establishes the dense-relationship between audio and visual modalities. Specifically, the DIMA module leverages the information of all channel pairs of audio and visual features to formulate the cross-modality attention weight, which is superior to the multi-head attention module that uses limited information. Moreover, a novel unimodal discrimination loss (UDL) is introduced to exploit the unimodal and fused features together for more exact AVE localization. The experimental results show that our method is remarkably superior to the state-of-the-art methods in fully- and weakly-supervised AVE settings. To further evaluate the model's ability to build audio-visual connections, we design a dense cross modality relation network (DCMR) to solve the cross-modality localization task. DCMR is a simple deformation of a DMIN, and the experimental results further illustrate that DIMA can explore denser relationships between the two modalities. Code is available at https://github.com/weizequan/DMIN.git.

In this paper we explore automatic genre labelling of motion picture previews using audio-visual features present in movie trailers and the focus is on fusion techniques (early fusion and late fusion) and the resultant improvement on classification accuracy. This paper proposes a novel combination of deep learned features (from a pretrained VGG-16 model) obtained using a state-of-the-art shot detector and hand-crafted audio features. This combination of features and an associated comparison of early and late fusion with these features has not been attempted in the literature before. Furthermore, two popular fusion techniques and three distinct classification algorithms are investigated to determine the optimal fusion technique and classifier combination. The study uses a subset of the LMTD-9 movie trailer dataset with selected genres (action, comedy, drama and horror). The best performing low-level audio features are comprised of timbre features extracted using the MIRtoolbox followed by standalone mel-frequency cepstral coefficients. The best performing high-level audio feature is tonality. Audio features are augmented by visual features extracted using a pre-trained convolutional neural network (VGG-16). Feature fusion (early and late fusion) methods are investigated together with classification methods such as extreme gradient boosting, support vector machine and a neural network. Evaluation metrics such as precision, recall, confusion matrices and F1 score are used to measure classification accuracy. Early fusion methods outperform late fusion methods with a classification performance gain of approximately 10% for a four class classification problem. The best classification performance for early fusion obtained with a support vector machine is (73.12% accuracy), followed by the extreme gradient boosting classifier (69.37% accuracy) and neural network classifier (67.50% accuracy), whereas chance is 25%. It is shown that superior classification performance can be achieved by employing early feature fusion of low-level audio descriptors, high-level audio descriptors and high-level visual feature descriptors together with suitable classifiers.

The aim of this paper is to categorize movies into genres using the previews. Our study attempts to combine audio, visual and text features to classify a collection of movie previews into action, biography, comedy, and horror. For each of the collected previews, the audio and visual features are extracted and the text features are drawn from social tags via social websites. The probabilistic latent semantic analysis (PLSA) is used to incorporate the features from these three different aspects of information. The standard PLSA processes one type of information only. Therefore double-model and triple-model PLSAs are extended in order to combine two or three different types of information. We compare these various variants of PLSA approaches with unimodal PLSAs, which use either audio, visual or text features only. The experimental results show not only that one of the triple-model PLSAs achieves the highest accuracy, but also that social tags (text features) play an important role for classifying movies genres.

This paper proposes an audio-visual speech recognition method using lip information extracted from side-face images as an attempt to increase noise robustness in mobile environments. Our proposed method assumes that lip images can be captured using a small camera installed in a handset. Two different kinds of lip features, lip-contour geometric features and lip-motion velocity features, are used individually or jointly, in combination with audio features. Phoneme HMMs modeling the audio and visual features are built based on the multistream HMM technique. Experiments conducted using Japanese connected digit speech contaminated with white noise in various SNR conditions show effectiveness of the proposed method. Recognition accuracy is improved by using the visual information in all SNR conditions. These visual features were confirmed to be effective even when the audio HMM was adapted to noise by the MLLR method.

An audio-visual speech recognition system (AVSR) integrates audio and visual information to perform speech recognition task. The AVSR has various applications in practice especially in natural language processing systems such as speech-to-text conversion, automatic translation and sentiment analysis. Decades ago, researchers tend to use Hidden Markov Model (HMM) to construct speech recognition system due to its good achievements in success recognition rate. However, HMM's training dataset is enormous in order to have sufficient linguistic coverage. Besides, its recognition rate under noisy environments is not satisfying. To overcome this deficiency, a Recurrent Neural Network (RNN) based AVSR is proposed. The proposed AVSR model consists of three components: 1) audio features extraction mechanism, 2) visual features extraction mechanism and 3) audio and visual features integration mechanism. The features integration mechanism combines the output features from both audio and visual extraction mechanisms to generate final classification results. In this research, the audio features mechanism is modelled by Mel-frequency Cepstrum Coefficient (MFCC) and further processed by RNN system, whereas the visual features mechanism is modelled by Haar-Cascade Detection with OpenCV and again, it is further processed by RNN system. Then, both of these extracted features were integrated by multimodal RNN-based features-integration mechanism. The performance in terms of the speech recognition rate and the robustness of the proposed AVSR system were evaluated using speech under clean environment and Signal Noise Ratio (SNR) levels ranging from -20 dB to 20 dB with 5 dB interval. On average, final speech recognition rate is 89% across different levels of SNR.

The Dense Audio-Visual Event Localization (DAVEL) task aims to temporally localize events in untrimmed videos that occur simultaneously in both the audio and visual modalities. This paper explores DAVEL under a new and more challenging weakly-supervised setting (W-DAVEL task), where only video-level event labels are provided and the temporal boundaries of each event are unknown. We address W-DAVEL by exploiting cross-modal salient anchors, which are defined as reliable timestamps that are well predicted under weak supervision and exhibit highly consistent event semantics across audio and visual modalities. Specifically, we propose a Mutual Event Agreement Evaluation module, which generates an agreement score by measuring the discrepancy between the predicted audio and visual event classes. Then, the agreement score is utilized in a Cross-modal Salient Anchor Identification module, which identifies the audio and visual anchor features through global-video and local temporal window identification mechanisms. The anchor features after multimodal integration are fed into an Anchor-based Temporal Propagation module to enhance event semantic encoding in the original temporal audio and visual features, facilitating better temporal localization under weak supervision. We establish benchmarks for W-DAVEL on both the UnAV-100 and ActivityNet1.3 datasets. Extensive experiments demonstrate that our method achieves state-of-the-art performance.

One of the most commonly used audiovisual fusion approaches is feature-level fusion where the audio and visual features are concatenated. Although this approach has been successfully used in several applications, it does not take into account interactions between the features, which can be a problem when one and/or both modalities have noisy features. In this paper, we investigate whether feature fusion based on explicit modelling of interactions between audio and visual features can enhance the performance of the classifier that performs feature fusion using simple concatenation of the audio-visual features. To this end, we propose a log-linear model, named Bimodal Log-linear regression, which accounts for interactions between the features of the two modalities. The performance of the target classifiers is measured in the task of laughter-vs-speech discrimination, since both laughter and speech are naturally audiovisual events. Our experiments on the MAHNOB laughter database suggest that feature fusion based on explicit modelling of interactions between the audio-visual features leads to an improvement of 3\% over the standard feature concatenation approach, when log-linear model is used as the base classifier. Finally, the most and least influential features can be easily identified by observing their interactions.

This paper aims at comparing the discrimination between audio, 2D-based visual and 3D-based visual features for the speech recognition purpose. The audio and visual feature extraction schemes and several feature selection techniques are described first in this paper. With the application of the described feature extraction and selection methods, several experiments are conducted to compare the discrimination of the audio features, the 2D visual features and the 3D visual features for the hVd words classification task. In our study, it is found that the 3D visual features have more separability than the 2D visual features, so that the 3D-based audio-visual speech recognition may achieve more desirable results than the traditional 2D-based counterpart.

Automatic recognition of eating conditions of humans could be a useful technology in health monitoring. The audio-visual information can be used in automating this process, and feature engineering approaches can reduce the dimensionality of audio-visual information. The reduced dimensionality of data (particularly feature subset selection) can assist in designing a system for eating conditions recognition with lower power, cost, memory and computation resources than a system which is designed using full dimensions of data. This paper presents Active Feature Transformation (AFT) and Active Feature Selection (AFS) methods, and applies them to all three tasks of the ICMI 2018 EAT Challenge for recognition of user eating conditions using audio and visual features. The AFT method is used for the transformation of the Mel-frequency Cepstral Coefficient and ComParE features for the classification task, while the AFS method helps in selecting a feature subset. Transformation by Principal Component Analysis (PCA) is also used for comparison. We find feature subsets of audio features using the AFS method (422 for Food Type, 104 for Likability and 68 for Difficulty out of 988 features) which provide better results than the full feature set. Our results show that AFS outperforms PCA and AFT in terms of accuracy for the recognition of user eating conditions using audio features. The AFT of visual features (facial landmarks) provides less accurate results than the AFS and AFT sets of audio features. However, the weighted score fusion of all the feature set improves the results.

Chapter 21 - Optical flow-based representation for video action detection

In this paper, a new approach to visual speech synthesis using a joint probabilistic model is introduced, namely the Gaussian process latent variable model trimmed with manifold relevance determination model, which explicitly models coarticulation. One talking head dataset is processed (LIPS dataset) by extracting visual and audio features from the sequences. The model can capture the structure of data with extremely high dimensionality. Distinguishable visual features can be inferred directly from the trained model by sampling from the discovered latent points. Statistical evaluation of inferred visual features against ground truth data is obtained and compared with the current state-of-the-art visual speech synthesis approach. The quantitative results demonstrate that the proposed approach outperforms the state-of-the-art technique.

In this paper a fast and robust method is proposed to search a large video collection for given short clips. Compared with existing video searching methods which use visual features only, our scheme performs a two-phase hierarchical matching technique using visual and audio features successively. Considering that video sampling rate (25 or 30 fps) is much lower than that of audio (8 to 48 kHz), a coarse search is implemented with sub-sampled video frames first, and then potential matches are verified and accurately located using fine audio features. Both features are extracted directly from MPEG compressed video for computational efficiency. Experiments have been conducted on over 10.5 hours of video to search for re-occurrences of 83 TV commercials and one news lead-out clip. All the 220 instances are correctly detected with no false alarm. Our experiments also show that the proposed method is robust to variations of video bit rate, frame rate, frame size and color shifting.

Sound source localization aims to localize objects emitting the sound in visual scenes. Recent works obtaining impressive results typically rely on contrastive learning. However, the common practice of randomly sampling negatives in prior arts can lead to the false negative issue, where the sounds semantically similar to visual instance are sampled as negatives and incorrectly pushed away from the visual anchor/query. As a result, this misalignment of audio and visual features could yield inferior performance. To address this issue, we propose a novel audio-visual learning framework which is instantiated with two individual learning schemes: self-supervised predictive learning (SSPL) and semantic-aware contrastive learning (SACL). SSPL explores image-audio positive pairs alone to discover semantically coherent similarities between audio and visual features, while a predictive coding module for feature alignment is introduced to facilitate the positive-only learning. In this regard SSPL acts as a negative-free method to eliminate false negatives. By contrast, SACL is designed to compact visual features and remove false negatives, providing reliable visual anchor and audio negatives for contrast. Different from SSPL, SACL releases the potential of audio-visual contrastive learning, offering an effective alternative to achieve the same goal. Comprehensive experiments demonstrate the superiority of our approach over the state-of-the-arts. Furthermore, we highlight the versatility of the learned representation by extending the approach to audio-visual event classification and object detection tasks.