Search

This paper proposes a novel method to utilize three source features for video captioning. It fuses global video features with local object and regional features to model the relationships among objects and their motions and applies object tags instead of visual features to guide the generation of descriptions. Specifically, Multi-feature is firstly extracted by pretrained models and treated as separate inputs alongside video frames. Secondly, an object awareness attention block is designed to fuse the different features information and to learn a joint video representation which has both visual and linguistic semantics. Experiments on MSVD and MSR-VTT datasets have shown the effectiveness of the proposed method, and the ablation studies have verified the contribution of each component.

Perceiving user emotions and generating responses with specific emotions are of great significance to a social dialogue system. Much previous work still utilizes seq2seq models as the backbone to generate emotional responses. With the popularity of the generative pre-training models, we propose an emotional response generation model EmoDialoGPT based on DialoGPT by introducing emotion embedding and emotion prediction loss. In order to obtain a large-scale dialogue dataset with emotion labels, we train an emotion classifier and automatically annotate the emotion labels for the dialogue data. We evaluate our models from three aspects, including emotion expression, response quality and human evaluation. The experimental results show that, the proposed models outperform seq2seq models and baseline DialoGPT models on most metrics.

BitTorrent was long believed to be robust to free-riding behavior. Yet, some recent studies show that misbehaving peers download far more than they upload and free riders even upload nothing at all. In order to guarantee the upload fairness among peers, this paper proposes a quota-based encrypted block trading protocol (Quota-Encryption for short). The basic idea is for peers to prepay a ''quota'' of encrypted blocks to each other and then to alternately reveal decryption keys. Simulation results show that it guarantees better upload fairness among peers, making the free-riding behavior impossible in BitTorrent.

We consider the problem of license plate detection in natural scenes using Convolutional Neural Network (CNN). CNNs are global trainable multi-stage architectures that automatically learn shift invariant features from the raw input images. Additionally, they can be easily replicated over the full input making them widely used for object detection. However, such detectors are currently limited to single-scale architecture in which the classifier only use the features extracted by last stage. In this paper, a multi-scale CNN architecture is proposed in which the features extracted by multiple stages are fed to the classifier. Furthermore, additional subsampling layers are added making the presented architecture also easily replicated over the full input. We apply the proposed architecture to detect license plates in natural sense images, and it achieves encouraging detection rate with neither handcrafted features nor controlling the image capturing process.

Video-on-Demand (VoD) is popular on the Internet for offering time-unconstrained viewing experience. In order to reduce their operational costs, VoD service providers call for assistance of the participating peers via employing peer-to-peer technologies. However, perfect viewing experiences perceived on the home DVD players are still hardly (or impractical) implemented on the Internet because of Internet features, provider costs, media characteristics, and viewer behaviors. In this paper, we are motivated to take viewer behaviors into account. Specifically, information of media chunks being available in neighborhood is compactly represented on the interface of viewers' clients. The viewers in turn may bias their future VCR operations towards the currently available chunks for achieving a seldom interrupted media playback. As a side effect, VoD providers are more likely to reduce their operational costs for serving asynchronously arriving viewers.

P2PTV refers to television media streams being distributed to large numbers of viewers on the Internet via peer-to-peer overlay networks. In the design of P2PTV systems, there is a fundamental trade-off between the better Quality-of-Service (QoS) desired by the viewers and the lower resource consumption preferred by the P2PTV distributors as well as the Internet Service Providers (ISPs). This trade-off is partially caused by the time-varying and limited capabilities of the overlay connections. Observing the potential duplication of the television media streams, we propose a metadata-based delivery framework for P2PTV. In contrast to previous efforts, this framework distributes the original media streams as well as the associated metadata information simultaneously. As a result, the streaming clients at the viewer side are able to identify duplicated content segments in media streams and avoid requesting the duplicated content multiple times. In this way, the P2PTV distributors, ISPs and viewers could reduce their upload bandwidth consumption. Moreover, the distributors within this framework can discriminate various media content according to time constraints such that they could apply different uploading policies to live, pre-recorded and hybrid media content. The streaming clients are even allowed to slightly time-shift local playback schedules such that viewers have an opportunity to experience better P2PTV services.

Peer-to-Peer (P2P) media streaming applications are increasingly popular on the Internet and offer a wide variety of media content, attracting potentially millions of viewers simultaneously. However, much content is duplicated, synchronously and asynchronously: popular Hollywood movies are often offered by many different on-demand streaming service providers; live streaming service providers distribute the same media content, for instance television series, in different channels over a delay of some hours or days. Viewers are more concerned about the media content and the quality of the streaming service, and less about who is actually offering it. In this article we aim at enriching the viewers' options instead of focusing on particular streaming applications: 1) we present a general content service facilitating the viewers to share their choices of what, when and where they will watch; 2) we extend streaming clients to accommodate various behaviors of the viewers besides their traditional operations. As such, the viewers will get a better experience, and streaming service providers and Internet service providers both have an opportunity to cut down their resource consumption.

Trust modeling is currently a key part in building P2P and other internet-based systems. One way for trust generation is to view the trust among users as a web of trust in large scale internet settings, and the trust quantities are further computed from such a web. However, due to the large size and sparseness of the web of trust, how to effectively represent and propagate trust in such a web remains a challenging problem. In this paper, we first present a novel way for similarity measurement from the point of view of user feedback patterns, which can be further incorporated into a trust model. And then the efficiency and effectiveness of the proposed method is given in simulations.

Collaborative filtering as a classical method of information retrieval is widely used in helping people to deal with information overload. In this paper, we introduce the concept of local user similarity and global user similarity, based on surprisal-based vector similarity and the application of the concept of maximin distance in graph theory. Surprisal-based vector similarity expresses the relationship between any two users based on the quantities of information (called surprisal) contained in their ratings, which is based on the intuition that less common ratings for a specific item tend to provide more discriminative information than the most common ratings. Furthermore, traditional methods of computing user similarity can not work if two users have not rated any identical item. To solve this problem, the global user similarity is introduced to define two users being similar if they can be connected through their locally similar neighbors. A weighted user graph is first constructed by using local similarity of any two users as the weight of the edge connecting them. Then the global similarity can be calculated as the maximin distance of any two nodes in the graph. Based on both of Local User Similarity and Global User Similarity, we develop a collaborative filtering framework called LS&GS. An empirical study using the MovieLens dataset shows that the proposed framework outperforms other state-of-the-art collaborative filtering algorithms.

Peer-to-peer technique becomes mature gradually, and multiple domain-involved applications emerge, such as IPTV, distance learning, chatting network. In the context of distance learning, small scale of teachers would serve large scale of geographically located students. Normally, knowledge points are associated with different difficulty levels. And students usually are interested in varied subsets of knowledge points. Also teachers are capable of serving knowledge point subsets. The objective to schedule meeting among students and teachers according to their respective interests and capabilities is to reduce total learning duration. After formulating meeting schedule as Integer Programming problem, this paper proposes three heuristic algorithms to approximate the optimal solution. To the best of our knowledge, such problem is firstly investigated in distance learning context. Performance evaluation demonstrates their behaviors and PKPA algorithm excels two others substantially.

Peer-to-peer overlay is self-organized by individual hosts in the Internet. The overlay usually serves for collaboration-oriented applications and necessitates group communication. Consequently, anonymity is demanded in scenarios such as military, government applications for privacy protection. It prevents communicating party identity from being revealed by others. In this paper we propose Bus-based anonymous multicast upon peer-to-peer overlay, which delivers message from originator to consumers anonymously. The Bus rotates over an Euler tour formed by hosts and takes intentional messages to respective consumers. Theoretical analysis demonstrates that bus-based anonymous multicast gets higher anonymity level comparing to existing ones.

Data-driven media streaming has been deployed gradually over the internet. in such systems, node periodically exchanges media block availability and fetches desirable blocks from neighbors. The issue on optimizing fetching blocks is called block scheduling and receives focus, especially in heterogeneous overlay. in this paper Bandwidth-Aware Scheduling (BAS) formulates such problem by incorporating bandwidth and block deadline, as well as node importance. Simulation results verify that BAS performs much better than existing ones.

Recently directional sensors have been widely deployed as they are more practicable under constraints of manufacture, size and cost. One common functionality of networks formed by such directional sensors is to monitor a set of discrete targets continuously. Large scale deployment makes sensor recharge impossible. By abundant deployment, it is reasonable and necessary to select subsets of sensors to operate alternatively so as to prolong the network lifetime. Such problem has been proved to be NP-Complete. This paper approximates network lifetime problem by randomized algorithm. Through constructing elementary sessions, which denotes active subset of sensors covering all targets, and linear programming, the approximating solution is derived within extremely less duration comparing to previous works. Simulation results demonstrate the algorithm's performance and sound explanation is also presented.

Peer-to-peer networks can be divided into structured and unstructured based on their overlay topologies. In reality, unstructured p2p networks with superpeers have proved their capacities to support millions of users simultaneously. However, applications deployed on this kind of overlay networks, such as file-sharing, require flooding or gossip-based message routing, which puts more overhead on underlying networks and provides no guarantee on resource discovery. In this paper we propose an overlay adaptation protocol which allows structured superpeer overlay construction from unstructured p2p overlay networks with the potential to leverage the advantages of structured p2p overlay networks such as efficiency, scalability and guaranteed look-up services. The simulation result shows that our protocol can build the structured superpeer overlay with efficiency and scalability.

Sensor networks have been applied in a wide variety of situations. Recently directional sensor networks consisting of directional sensors have gained attention. As for the traditional target coverage problem, the limited sensing angle of directional sensors makes it even more challenging. Moreover, individual targets may also be associated with differentiated priorities. Considering the distance between the directional sensors and targets influences sensing quality, this paper proposes the priority-based target coverage problem and strives to choose a minimum subset of directional sensors that can monitor all targets, satisfying their prescribed priorities. Due to the NP-Complete complexity, the minimum subset of directional sensors is approximated by using a genetic algorithm. Simulation results reveal the effects of multiple factors on the size of the resulting subset.