一种基于多尺度实例学习的视频事件检测算法

现有的大多数视频事件检测方法首先从视频帧或视频快照中提取特征,然后对特征进行量化和汇集,进而为整个视频生成一个向量表示.最后的汇集步骤虽然简单高效,但是可能丢失时间局部信息,而这些信息对于确定长视频中事件发生的位置具有重要作用,从而削弱了事件检测的准确性.为此,本文首先将每个视频表示为多个“实例”,并将其定义为不同时间间隔的视频段.然后,针对每个视频的正实例比例已知和未知两种情况,提出基于多尺度实例学习的检测算法,在将实例标签看成隐藏潜在变量的同时推断出实例标签以及实例尺度的事件检测模型.最后,利用大规模视频事件数据集进行了全面的仿真实验,结果证明了本文算法具有显著的性能提升.此外,算法还可以确定视频中导致正检测的时间段的位置,进而对检测结果做出解释.

Video event detection algorithm based on multi-scale instance learning

The existing most video event detection approaches firstly extract features from video trames or shots,then quantize and pool the features to form a single vector representation for the entire video.Though simple and efficient,the final pooling step may lead to loss of temporally local information,which is important in indicating which part in a long video signifies presence of the event,and weak the accuracy of event detection.To this end,an instance-based video event detection approach is proposed.Each video is firstly represented as multiple "instances",which is defined as video segments of different temporal intervals.Then,aiming at the two cases of the proportion of the positive examples of each video,a detection algorithm based on multi-scale is proposed which treats the instance labels as hidden latent variables,and simultaneously infers the instance labels as well as the instance-level event detection model.Finally,extensive experiments on large-scale video event datasets demonstrate significant performance gains.In addition,the proposed method is also useful in explaining the detection results by localizing the temporal segments in a video which is responsible for the positive detection.