摘要 :
Researchers have investigated various graph embedding methods to complete Knowledge Graphs (KGs), most of which merely focus on Static KGs (SKGs) without emphasizing the time dependence of triple-formed facts. However, in reality,...
展开
Researchers have investigated various graph embedding methods to complete Knowledge Graphs (KGs), most of which merely focus on Static KGs (SKGs) without emphasizing the time dependence of triple-formed facts. However, in reality, KGs are dynamic and definitely there is correlations between facts with different timestamps. Due to the sparsity of Temporal KGs (TKGs), SKG's embedding methods cannot be directly applied to TKGs, which triggers the current discussions about TKG Completion (TKGC) task. And existing TKGC methods universally suffer from two issues: (ⅰ) The modeling procedure for temporal information in encoder is usually disjointed or conflict with that in decoder, (ⅱ) Current methods are overwhelmingly dependent on temporal signals for measuring the probability of candidate entity, while ignoring other signals (such as entity's semantics, etc.,). To overcome these problems, this paper proposes a novel semantic-driven time-aware relational graph neural network model for TKGC task, which consists of a semantic-enhanced encoder and a convolution-based decoder.
收起
摘要 :
Nowadays, digital videos have been widely leveraged to record and share various events and people's daily life. It becomes urgent to provide automatic video semantic analysis and management for convenience. Shot boundary detection...
展开
Nowadays, digital videos have been widely leveraged to record and share various events and people's daily life. It becomes urgent to provide automatic video semantic analysis and management for convenience. Shot boundary detection (SBD) plays a key fundamental role in various video analysis. Shot boundary detection aims to automatically detecting boundary frames of shots in videos. In this paper, we propose a progressive method for shot boundary detecting with histogram based shot filtering and C3D based gradual shot detection. Abrupt shots were detected firstly for its specialty and help alleviate locating shots across different shots by dividing the whole video into segments. Then, over the segments, gradual shot detection is implemented via a three-dimensional convolutional neural network model, which assign video clips into shot types of normal, dissolve, foi or swipe. Finally, for untrimmed videos, a frame level merging strategy is constructed to help locate the boundary of shots from neighboring frames. The experimental results demonstrate that the proposed method can effectively detect shots and locate their boundaries.
收起
摘要 :
Nowadays, digital videos have been widely leveraged to record and share various events and people's daily life. It becomes urgent to provide automatic video semantic analysis and management for convenience. Shot boundary detection...
展开
Nowadays, digital videos have been widely leveraged to record and share various events and people's daily life. It becomes urgent to provide automatic video semantic analysis and management for convenience. Shot boundary detection (SBD) plays a key fundamental role in various video analysis. Shot boundary detection aims to automatically detecting boundary frames of shots in videos. In this paper, we propose a progressive method for shot boundary detecting with histogram based shot filtering and C3D based gradual shot detection. Abrupt shots were detected firstly for its specialty and help alleviate locating shots across different shots by dividing the whole video into segments. Then, over the segments, gradual shot detection is implemented via a three-dimensional convolutional neural network model, which assign video clips into shot types of normal, dissolve, foi or swipe. Finally, for untrimmed videos, a frame level merging strategy is constructed to help locate the boundary of shots from neighboring frames. The experimental results demonstrate that the proposed method can effectively detect shots and locate their boundaries.
收起
摘要 :
The core of recommendation systems is to explore users' preferences from users' historical records and accordingly recommend items to meet users' interests. Previous works explore interaction graph to capture multi-order collabora...
展开
The core of recommendation systems is to explore users' preferences from users' historical records and accordingly recommend items to meet users' interests. Previous works explore interaction graph to capture multi-order collaborative signals and derive high-quality representations of users and items, which effectively alleviates the interaction spar-sity issue. Recent works extend the scope with a fine-grained perspective and achieve a great success in modeling users' diverse intents. Although these works distinguish intents, they ignore the hidden correlation among users' intents resulting in suboptimal recommendation performance. We argue that a user's interest is made up of multiple intents and these intents are compatible on the interest composition of the user. To this point, we propose multi-intent compatible transformer network (MCTN) to explore the correlation between intents on modeling users' interests for recommendation. Users and items are embedded into multiple intent spaces through disentangled graph convolution network to disentangle users' intents. MCTN conducts embedding propagation in each intent space to capture the multi-order collaborative signals on the specific intent. We introduce a transformer network to capture the dependence between intents and derive multi-intent compatible embeddings of users and items for recommendation. The experiments achieves state-of-the-art performance, which demonstrates the effectiveness of the proposed MCTN on modeling multi-intent compatibility into embeddings.
收起
摘要 :
Personalized recommendation refers to identifying items that satisfy users' interests from large-scale item databases according to users' habits and preferences. The task is very challenging due to the complexity of user interests...
展开
Personalized recommendation refers to identifying items that satisfy users' interests from large-scale item databases according to users' habits and preferences. The task is very challenging due to the complexity of user interests. Previous works use a uniform representation to model user interests, neglecting the diversity of user preferences when they adopt items. However, users consider many different attributes when choosing an item. Introducing attribute-level matching information into the model can express user interests more accurately. To achieve this goal, we propose a novel Attribute-level Interest Matching Network (AIMN) for personalized recommendation. We first adopt a knowledge representation learning method to construct spaces of different attributes, then employ a knowledge graph to extend entities as side information for representing users. Finally, we project entities and candidate items into diverse attribute spaces, match and aggregate them to realize fine-grained attribute-level information matching. Empirical results demonstrate that the proposed AIMN achieves substantial gains on several benchmarks, beating many solid baselines and achieving state-of-art performance.
收起
摘要 :
Personalized recommendation refers to identifying items that satisfy users' interests from large-scale item databases according to users' habits and preferences. The task is very challenging due to the complexity of user interests...
展开
Personalized recommendation refers to identifying items that satisfy users' interests from large-scale item databases according to users' habits and preferences. The task is very challenging due to the complexity of user interests. Previous works use a uniform representation to model user interests, neglecting the diversity of user preferences when they adopt items. However, users consider many different attributes when choosing an item. Introducing attribute-level matching information into the model can express user interests more accurately. To achieve this goal, we propose a novel Attribute-level Interest Matching Network (AIMN) for personalized recommendation. We first adopt a knowledge representation learning method to construct spaces of different attributes, then employ a knowledge graph to extend entities as side information for representing users. Finally, we project entities and candidate items into diverse attribute spaces, match and aggregate them to realize fine-grained attribute-level information matching. Empirical results demonstrate that the proposed AIMN achieves substantial gains on several benchmarks, beating many solid baselines and achieving state-of-art performance.
收起
摘要 :
Multiple object tracking (MOT) plays a key role in video analysis. On MOT, DeepSORT (Simple Online and Realtime Tracking with a deep association metric) performs effectively by combining features of appearance and motion for estim...
展开
Multiple object tracking (MOT) plays a key role in video analysis. On MOT, DeepSORT (Simple Online and Realtime Tracking with a deep association metric) performs effectively by combining features of appearance and motion for estimating data association. However, computing with multiple features are time consuming. In certain applications, cameras are static, such as pedestrian surveillance, sports video analysis and so on. Here, without camera movement the motion trajectories of objects are generally possible to estimate. The introduction of more features cannot improve the performance of object tracking discriminatively. Furthermore, the time cost rises evidently. To address this problem, we propose a novel Simple Online and Realtime Tracking with motion features (MF-SORT). By focusing on the motion features of the objects during data association, the proposed scheme is able to take a trade-off between performance and efficiency. The experimental results on the MOT Challenge benchmark and MOT-SOCCER (newly established in this work) demonstrate that the proposed method is much faster than DeepSORT with the comparable accuracy.
收起
摘要 :
Deep learning methods show strong ability in extracting high-level features for images in the field of person re-identification. The produced features help inherently distinguish pedestrian identities in images. However, on deep l...
展开
Deep learning methods show strong ability in extracting high-level features for images in the field of person re-identification. The produced features help inherently distinguish pedestrian identities in images. However, on deep learning models over-fitting and discriminative ability of the learnt features are still challenges for person re-identification. To alleviate model over-fitting and further enhance the discriminative ability of the learnt features, we propose Siamese pedestrian alignment networks (SPAN) for person re-identification. SPAN employs two streams of PAN (pedestrian alignment networks) to increase the size of network inputs over limited training samples and effectively alleviate network over-fitting in learning. In addition, a verification loss is constructed between the two PANs to adjust the relative distance of two input pedestrians of the same or different identities in the learned feature space. Experimental verification is conducted on six large person re-identification datasets and the experimental results demonstrate the effectiveness of the proposed SPAN for person re-identification.
收起
摘要 :
Deep learning methods show strong ability in extracting high-level features for images in the field of person re-identification. The produced features help inherently distinguish pedestrian identities in images. However, on deep l...
展开
Deep learning methods show strong ability in extracting high-level features for images in the field of person re-identification. The produced features help inherently distinguish pedestrian identities in images. However, on deep learning models over-fitting and discriminative ability of the learnt features are still challenges for person re-identification. To alleviate model over-fitting and further enhance the discriminative ability of the learnt features, we propose Siamese pedestrian alignment networks (SPAN) for person re-identification. SPAN employs two streams of PAN (pedestrian alignment networks) to increase the size of network inputs over limited training samples and effectively alleviate network over-fitting in learning. In addition, a verification loss is constructed between the two PANs to adjust the relative distance of two input pedestrians of the same or different identities in the learned feature space. Experimental verification is conducted on six large person re-identification datasets and the experimental results demonstrate the effectiveness of the proposed SPAN for person re-identification.
收起
摘要 :
Multiple object tracking (MOT) plays a key role in video analysis. On MOT, DeepSORT (Simple Online and Realtime Tracking with a deep association metric) performs effectively by combining features of appearance and motion for estim...
展开
Multiple object tracking (MOT) plays a key role in video analysis. On MOT, DeepSORT (Simple Online and Realtime Tracking with a deep association metric) performs effectively by combining features of appearance and motion for estimating data association. However, computing with multiple features are time consuming. In certain applications, cameras are static, such as pedestrian surveillance, sports video analysis and so on. Here, without camera movement the motion trajectories of objects are generally possible to estimate. The introduction of more features cannot improve the performance of object tracking discriminatively. Furthermore, the time cost rises evidently. To address this problem, we propose a novel Simple Online and Realtime Tracking with motion features (MF-SORT). By focusing on the motion features of the objects during data association, the proposed scheme is able to take a trade-off between performance and efficiency. The experimental results on the MOT Challenge benchmark and MOT-SOCCER (newly established in this work) demonstrate that the proposed method is much faster than DeepSORT with the comparable accuracy.
收起