Boundary Sensitive and Confidence Fusion Network for Temporal Action Localization

摘要

In long and untrimmed videos with rich contents, action localization has become a big challenge for both detection tasks and contents analyses. Nowadays, temporal action detection algorithms based on deep learning have shown good performance via making a full use of temporal information. These algorithms still have the potential to get higher recall in small-scale action instances, since the confidence scores may lack of representation and accuracy. To address these difficulties in temporal action proposal stage, we present an effective confidence fusion method, called Boundary Sensitive and Confidence Fusion network (BSCF), which fuses the confidence scores with a stream of optical flow, based on Boundary-Sensitive Network (BSN). Experiments show that the area under AR-AN curve of BSCF on ActivityNet-v3 dataset is larger than BSN by 0.43%.