Temporal attention fusion for sports event detection
Reede Ren, Yue Jiang Feng, Joemon M. Jose · 2008
Keywords:attention fusion, event detection, video retrieval The employment of psychological measurement,attention, alleviates the semantic uncertainty around video events and leads to an effective general event detection approach. This paper proposes a multi-resolution autoregressive framework to estimate a unified attention curve from multi-modality salient features at different temporal resolutions. The highlights of this work are: (1) the capability of using data at very coarse temporal resolutions, e.g. three minutes; (2) the robustness against noise caused by modality asynchronism and feature collection size; and (3) the utilisation of Markovian temporal constrains on content presentation. This approach achieved 100 % goal event coverage in the football video collection of the FIFA World Cup 2002, 2006 and UEFA League 2006. An attention perception system [11] consists of three components (Figure 1): pre-attentive, attention combination and post-attentive system. The pre-attentive system quickly calculates stimulus strength or extracts salient features. Additionally, such a process is also mentioned as featureattention modelling [7]. However, we do not assume this extraction is complete or even that these features can distinguish possible attention peaks from the background noise by themselves. This is because of strong noise in perception. The attention combination is to simulate attention mechanism in human minds. In this stage, salient features are fused as stimuli from vision, auditory and text understanding, to estimate a unified attention reflection. The post-attentive system justifies conclusions of attention combination by prior knowledge of video contents and thus completes event segmentation. 1