Improved action proposals using fine-grained proposal features with recurrent attention models

Recent models for the temporal action proposal task show that local properties can be an alternative to region network (RPN) generating good candidates on untrimmed videos. In this study, we devise RPN model with a new two-stage pipeline and joint scoring function proposals. The evaluation of is integrated into our search best distinguished mainly in fine details regions. Our proposals multiple scales using two recurrent neural layers attention mechanisms. We observe training clues multi-scale modeling mechanisms improve performance generation task. yields state-of-the-art results THUMOS-14 comparable ActivityNet-1.3 datasets.