Video Activity Localisation with Uncertainties in Temporal Boundary

Current methods for video activity localisation over time assume implicitly that temporal boundaries labelled model training are determined and precise. However, in unscripted natural videos, different activities mostly transit smoothly, so it is intrinsically ambiguous to determine labelling precisely when an starts ends time. Such uncertainties currently ignored training, resulting learning mis-matched video-text correlation with poor generalisation test. In this work, we solve problem by introducing Elastic Moment Bounding (EMB) accommodate flexible adaptive towards modelling universally interpretable tolerance underlying pre-fixed annotations. Specifically, construct elastic adaptively mining discovering frame-wise endpoints can maximise the alignment between segments query sentences. To enable both more accurate matching (segment content attention) robust boundaries), optimise selection of subject segment-wise contents a novel Guided Attention mechanism. Extensive experiments on three benchmarks demonstrate compellingly EMB’s advantages existing without uncertainty.