Optimization of Action Recognition Model Based on Multi-Task Learning and Boundary Gradient

Feature (linguistics)
DOI: 10.3390/electronics10192380 Publication Date: 2021-09-29T12:27:44Z
ABSTRACT
Recently, people’s demand for action recognition has extended from the initial high classification accuracy to of temporal detection. It is challenging meet two requirements simultaneously. The key behavior lies in quantity and quality extracted features. In this paper, a two-stream convolutional network used. A three-dimensional neural (3D-CNN) used extract spatiotemporal features consecutive frames. two-dimensional (2D-CNN) spatial key-frames. integration networks excellent improving model’s can complete task distinguishing start–stop frame. multi-scale feature extraction method presented more abundant information. At same time, multi-task learning model introduced. further improve via sharing data between multiple tasks. experimental result shows that modified improved by 10%. Meanwhile, we propose confidence gradient, which optimize frame detection accuracy. been enhanced 11%.
SUPPLEMENTAL MATERIAL
Coming soon ....
REFERENCES (29)
CITATIONS (4)