From: A multisource fusion framework driven by user-defined knowledge for egocentric activity recognition
Proposed method | Method in [15] | Method in [16] | |
---|---|---|---|
Vision | One pre-trained CNN (single image) + entropy-based TF-IDF | Three-stream CNN (single frame, optical flow, and stabilized optical flow) | Optical flow-based dense trajectory |
Low | High | Very high | |
Sensors | SVM | Four-stream LSTM | Temporal enhanced trajectory-like features |
Low | Medium | Low | |
Fusion | DSmT | Average or maximum pooling | Multimodal Fisher vector |
Medium | Low | Medium |