Weak supervision time sequence action positioning method based on adaptive time sequence feature enhancement

• Main Authors: YANG JIANTAO, LIU SHENG, PAN SONGQI, FENG YUAN, TIAN XIAOPENG, ZHANG YINENG
• Format: Patent
• Language: Chinese; English
• Published: 09.01.2024
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention belongs to the technical field of time sequence action positioning, and discloses a weak supervision time sequence action positioning method based on adaptive time sequence feature enhancement, which realizes better action classification and positioning performance by performing time sequence enhancement on original RGB and optical flow features. It is worthy of notice that compared with a previous method, due to the deformable convolution dynamic receptive field, the method can adaptively learn the complex and changeable sequential relation inside the action and between the action and the background. In addition, a teacher model is introduced to better utilize and learn time sequence information, and modal consensus training is guided. According to the method, the precision of a weak supervision time sequence action positioning task is improved, and superior performance is achieved on two popular data sets compared with a recent method. 本发明属于时序动作定位技术领域，公开了一种基于自适应时序特征增强的弱监督时序动作定位方法，通过对原始RGB和光流特征