Fig. 4From: Pedestrian detection with motion features via two-stream ConvNetsTemporal differences in UCF-101 [21], activity recognition dataset we used to train temporal streamBack to article page