This paper addresses the Multi-Athlete Tracking (MAT) problem, which plays a crucial role in sports video analysis. There exist specific challenges in MAT, e.g., athletes share a high similarity in appearance and frequently occlude with each other, making existing approaches not applicable for this task. To address this problem, we propose a novel online multiple athlete tracking approach which make use of long-term temporal pose dynamics for better distinguishing different athletes. Firstly, we design a Pose-based Triple Stream Network (PTSN) based on Long Short-Term Memory (LSTM) networks, capable of modeling long-term temporal pose dynamics of athletes, including pose-based appearance, motion and athletes’ interaction clues. Secondly, we propose a multi-state online matching algorithm based on bipartite graph matching and similarity scores produced by PTSN. It is robust to noisy detections and occlusions due to the reliable transitions of multiple detection states. We evaluate our method on the APIDIS, NCAA Basketball and VolleyTrack databases, and the experiment results demonstrate its effectiveness.

No abstract

No abstract

Attributed graphs, where nodes are associated with a rich set of attributes, have been widely used in various domains. Among all the nodes, those with patterns that deviate significantly from others are of particular interest. There are mainly two challenges for anomaly detection. For one thing, we often encounter large graphs with lots of nodes and attributes in the real-life scenario, which requires a scalable algorithm. For another, there are anomalies w.r.t. both the structure and attribute in a mixed manner. The algorithm should identify all of them simultaneously. State-of-art algorithms often fail in some respects. In this paper, we propose the scalable algorithm called MixedAD. Theoretical analysis is provided to prove its superiority. Extensive experiments are also conducted on both synthetic and real-life datasets. Specifically, the results show that MixedAD often achieves the F1 scores greater than those of others by at least 25% and runs at least 10 times faster than the others.

In recent years, Artificial Intelligence based disease diagnosis has drawn considerable attention both in academia and industry. In medical scenarios, a well-trained classifier can effectively detect a disease with sufficient features associating with medical tests. However, such features are not always readily available due to the high cost of time and money associating with medical tests. To address this, this study identifies the diagnostic strategy learning problem and proposes a novel framework consisting of three components to learn a diagnostic strategy with limited features. First, as we often encounter incomplete medical records of the patients, a sequence encoder is designed to encode any set of information in various sizes into fixed-length vectors. Second, taking the output of the encoder as the input, a feature selector based on reinforcement learning techniques is proposed to learn the best feature sequence for diagnosis. Finally, with the best feature sequence, an oracle classifier is used to give the final diagnosis. To evaluate the performance of the proposed method, experiments are conducted on nine real medical datasets. The results suggest that the proposed method is effective for providing personalized diagnostic strategies and makes better diagnoses with fewer features compared with existing methods.