With the growing number of videos in video sharing platforms, how to facilitate the searching and browsing of the user-generated video has attracted intense attention by multimedia community. To help people efficiently search and browse relevant videos, summaries of videos become important. The prior works in multimodal video summarization mainly explore visual and ASR tokens as two separate sources and struggle to fuse the multimodal information for generating the summaries. However, the time information inside videos is commonly ignored. In this paper, we find that it is important to leverage the timestamps to accurately incorporate multimodal signals for the task. We propose a Time-Aware Multimodal Transformer (TAMT) with a novel short-term order-sensitive attention mechanism. The attention mechanism can attend the inputs differently based on time difference to explore the time information inherent inside video more thoroughly. As such, TAMT can fuse the different modalities better for summarizing the videos. Experiments show that our proposed approach is effective and achieves the state-of-the-art performances on both YouCookII and open-domain How2 datasets.

In this paper, we present token labeling---a new training objective for training high-performance vision transformers (ViTs). Different from the standard training objective of ViTs that computes the classification loss on an additional trainable class token, our proposed one takes advantage of all the image patch tokens to compute the training loss in a dense manner. Specifically, token labeling reformulates the image classification problem into multiple token-level recognition problems and assigns each patch token with an individual location-specific supervision generated by a machine annotator. Experiments show that token labeling can clearly and consistently improve the performance of various ViT models across a wide spectrum. For a vision transformer with 26M learnable parameters serving as an example, with token labeling, the model can achieve 84.4% Top-1 accuracy on ImageNet. The result can be further increased to 86.4% by slightly scaling the model size up to 150M, delivering the minimal-sized model among previous models (250M+) reaching 86%. We also show that token labeling can clearly improve the generalization capability of the pretrained models on downstream tasks with dense prediction, such as semantic segmentation. Our code and model are publiclyavailable at https://github.com/zihangJiang/TokenLabeling.

Answering questions using multi-modal context is a challenging problem as it requires a deep integration of diverse data sources. Existing approaches only consider a subset of all possible interactions among data sources during one attention hop. In this paper, we present a Holistic Multi-modal Memory Network (HMMN) framework that fully considers interactions between different input sources (multi-modal context, question) at each hop. In addition, to hone in on relevant information, our framework takes answer choices into consideration during the context retrieval stage. Our HMMN framework effectively integrates information from the multi-modal context, question, and answer choices, enabling more informative context to be retrieved for question answering. Experimental results on the MovieQA and TVQA datasets validate the effectiveness of our HMMN framework. Extensive ablation studies show the importance of holistic reasoning and reveal the contributions of different attention strategies to model performance.

In this paper, we propose a novel unsupervised hashing method called Anchor-based Probability Hashing (APHash) to preserve the similarities by exploiting the distribution of data points. In particular, distances are transformed into probabilities in both original and hash code spaces. Our method aims to learn hash codes which minimize the mismatch between probability distributions of these two spaces. To address the high complexity issue, our method randomly selects a set of anchors and constructs asymmetric probability matrices. In this way, APHash can make use of the correlation between anchors and data points to learn hash codes more efficiently. Experimental results on two benchmark datasets demonstrate the effectiveness of the proposed APHash method, outperforming state-of-the-art hashing approaches in the application of image retrieval.

Hashing is an efficient approximate nearest neighbor search method and has been widely adopted for large-scale multimedia retrieval. While supervised learning is more popular for the data-dependent hashing, deep unsupervised hashing methods have recently been developed to learn non-linear transformations for converting multimedia inputs to binary codes. Most of existing deep unsupervised hashing methods make use of a quadratic constraint for minimizing the difference between the compact representations and the target binary codes, which inevitably causes severe information loss. In this paper, we propose a novel deep unsupervised method called DeepQuan for hashing. The DeepQuan model utilizes a deep autoencoder network, where the encoder is used to learn compact representations and the decoder is for manifold preservation. To contrast with the existing unsupervised methods, DeepQuan learns the binary codes by minimizing the quantization error through product quantization technique. Furthermore, a weighted triplet loss is proposed to avoid trivial solution and poor generalization. Extensive experimental results on standard datasets show that the proposed DeepQuan model outperforms the state-of-the-art unsupervised hashing methods for image retrieval tasks.

While convolutional neural networks (CNNs) have been excellent for object recognition, the greater spatial variability in scene images typically means that the standard full-image CNN features are suboptimal for scene classification. In this article, we investigate a framework allowing greater spatial flexibility, in which the Fisher vector (FV)-encoded distribution of local CNN features, obtained from a multitude of region proposals per image, is considered instead. The CNN features are computed from an augmented pixel-wise representation consisting of multiple modalities of RGB, HHA, and surface normals, as extracted from RGB-D data. More significantly, we make two postulates: (1) component sparsity—that only a small variety of region proposals and their corresponding FV GMM components contribute to scene discriminability, and (2) modal nonsparsity—that features from all modalities are encouraged to coexist. In our proposed feature fusion framework, these are implemented through regularization terms that apply group lasso to GMM components and exclusive group lasso across modalities. By learning and combining regressors for both proposal-based FV features and global CNN features, we are able to achieve state-of-the-art scene classification performance on the SUNRGBD Dataset and NYU Depth Dataset V2. Moreover, we further apply our feature fusion framework on an action recognition task to demonstrate that our framework can be generalized for other multimodal well-structured features. In particular, for action recognition, we enforce interpart sparsity to choose more discriminative body parts, and intermodal nonsparsity to make informative features from both appearance and motion modalities coexist. Experimental results on the JHMDB and MPII Cooking Datasets show that our feature fusion is also very effective for action recognition, achieving very competitive performance compared with the state of the art.

While convolutional neural networks (CNN) have been excellent for object recognition, the greater spatial variability in scene images typically meant that the standard full-image CNN features are suboptimal for scene classification. In this paper, we investigate a framework allowing greater spatial flexibility, in which the Fisher vector (FV) encoded distribution of local CNN features, obtained from a multitude of region proposals per image, is considered instead. The CNN features are computed from an augmented pixel-wise representation comprising multiple modalities of RGB, HHA and surface normals, as extracted from RGB-D data. More significantly, we make two postulates: (1) component sparsity - that only a small variety of region proposals and their corresponding FV GMM components contribute to scene discriminability, and (2) modal non-sparsity - within these discriminative components, all modalities have important contribution. In our framework, these are implemented through regularization terms applying group lasso to GMM components and exclusive group lasso across modalities. By learning and combining regressors for both proposal-based FV features and global CNN features, we were able to achieve state-of-the-art scene classification performance on the SUNRGBD Dataset and NYU Depth Dataset V2.

Most of the feature-learning methods for RGB-D object recognition either learn features from color and depth modalities separately, or simply treat RGB-D as undifferentiated four-channel data, which cannot adequately exploit the relationship between different modalities. Motivated by the intuition that different modalities should contain not only some modal-specific patterns but also some shared common patterns, we propose a multi-modal feature learning framework for RGB-D object recognition. We first construct deep CNN layers for color and depth separately, and then connect them with our carefully designed multi-modal layers, which fuse color and depth information by enforcing a common part to be shared by features of different modalities. In this way, we obtain features reflecting shared properties as well as modal-specific properties in different modalities. The information of the multi-modal learning frameworks is back-propagated to the early CNN layers. Experimental results show that our proposed multi-modal feature learning method outperforms state-of-the-art approaches on two widely used RGB-D object benchmark datasets.

Most existing approaches for RGB-D indoor scene labeling employ hand-crafted features for each modality independently and combine them in a heuristic manner. There has been some attempt on directly learning features from raw RGB-D data, but the performance is not satisfactory. In this paper, we propose an unsupervised joint feature learning and encoding (JFLE) framework for RGB-D scene labeling. The main novelty of our learning framework lies in the joint optimization of feature learning and feature encoding in a coherent way, which significantly boosts the performance. By stacking basic learning structure, higher level features are derived and combined with lower level features for better representing RGB-D data. Moreover, to explore the nonlinear intrinsic characteristic of data, we further propose a more general joint deep feature learning and encoding (JDFLE) framework that introduces the nonlinear mapping into JFLE. The experimental results on the benchmark NYU depth dataset show that our approaches achieve competitive performance, compared with the state-of-the-art methods, while our methods do not need complex feature handcrafting and feature combination and can be easily applied to other data sets.

Most existing feature learning-based methods for RGB-D object recognition either combine RGB and depth data in an undifferentiated manner from the outset, or learn features from color and depth separately, which do not adequately exploit different characteristics of the two modalities or utilize the shared relationship between the modalities. In this paper, we propose a general CNN-based multi-modal learning framework for RGB-D object recognition. We first construct deep CNN layers for color and depth separately, which are then connected with a carefully designed multi-modal layer. This layer is designed to not only discover the most discriminative features for each modality, but is also able to harness the complementary relationship between the two modalities. The results of the multi-modal layer are back-propagated to update parameters of the CNN layers, and the multi-modal feature learning and the back-propagation are iteratively performed until convergence. Experimental results on two widely used RGB-D object datasets show that our method for general multi-modal learning achieves comparable performance to state-of-the-art methods specifically designed for RGB-D data.

Spatial layout estimation of indoor rooms plays an important role in many visual analysis applications such as robotics and human-computer interaction. While many methods have been proposed for recovering spatial layout of rooms in recent years, their performance is still far from satisfactory due to high occlusion caused by the presence of objects that clutter the scene. In this paper, we propose a new approach to estimate the spatial layout of rooms from RGB-D videos. Unlike most existing methods which estimate the layout from still images, RGB-D videos provide more spatial-temporal and depth information, which are helpful to improve the estimation performance because more contextual information can be exploited in RGB-D videos. Given a RGB-D video, we first estimate the spatial layout of the scene in each single frame and compute the camera trajectory using the simultaneous localization and mapping (SLAM) algorithm. Then, the estimated spatial layouts of different frames are integrated to infer temporally consistent layouts of the room throughout the whole video. Our method is evaluated on the NYU RGB-D dataset, and the experimental results show the efficacy of the proposed approach.