Fig. 2From: A scale-adaptive object-tracking algorithm with occlusion detectionVisualization of the features from ResNet. a Input frame. b–e Features are extracted on convolutional layers conv1, conv2, conv3, and conv4 of ResNet; their original spatial resolutions are 128×128, 56×56, 28×28, and 14×14, respectively. The red boxes indicate the tracking results by our approachBack to article page