Improved Hierarchical Convolutional Features for Robust Visual Object Tracking

Figure 2

Visualization of convolutional layers on the challenging MotorRolling sequence. The yellow bounding boxes indicate the tracking results by our method. The red bounding boxes indicate the feature layer selected by our algorithm. Although the target is seriously deformed, the target position can still be identified by using the output features of the Conv5-4 layer. Compared with the features of the Conv5-pool layer, the resolution is too low to locate the accurate target position.