Weakly-supervised Temporal Action Localization

As social media gradually becomes the mainstream of society,network construction continues to improve,and video processing equipment is increasingly updated,the amount of video data reaches hundreds of millions and continues to grow at an explosive rate.The speed and accuracy of manual processing are no longer up to the requirements.For the huge amount of video data,it is necessary to use computer vision algorithms to process efficiently and accurately.Time sequence action positioning can be used in video surveillance,intelligent security,automatic extraction of sports videos,intelligent subtitle explanation and other scenarios,and the application fields are very wide.Therefore,temporal action localization has become an important task.The task requirement of time-series action localization is to detect the action start time,action end time and action category of the uncropped video.Although time-series action localization based on fully supervised learning has achieved good results,it requires video frame-level labels,which greatly increases labor and time costs.In contrast,time-series action localization based on weakly supervised learning only needs video-level labels to train a better model,and video-level labels do not require a lot of manual labeling and are easier to obtain attention of the industry.Aiming at the low accuracy of weakly supervised temporal action localization and the lack of video frame-level label supervision,this paper proposes a temporal action localization method based on dynamic thresholds and pseudo-supervision of attention and temporal action graphs simultaneously.In the weakly supervised time-series action localization method based on dynamic threshold,the dynamic threshold can dynamically control the number of binarized segments in the attention mechanism,and select the number of segments suitable for the video for different videos.A large number of experimental results and comparisons prove that the dynamic threshold method is superior to the state-of-the-art methods.The method of pseudo-supervision of attention value and time-series action graph at the same time makes the attention value and timeseries action graph that are not supervised by labels subject to the supervision of pseudo-true value best model.The experimental results exceed most state-of-the-art methods,showing that pseudo-supervision on both attention and temporal class action graphs can effectively improve model quality.