Figure 3.

Classification performance in variants of VGG-16 in comparison to humans. Human accuracies measured in a separate online experiment revealed that humans performed very well and virtually identical on photos and drawings and slightly worse on sketches. The ImageNet-trained VGG-16 (VGG-16 IN) showed high performance on photos but a sharp drop in performance for drawings and sketches. VGG-16 trained on stylized ImageNet (VGG-16 SIN) performed better on drawings than VGG-16 IN but not on sketches or photos. VGG-16 fine-tuned on ImageNet-Sketch (VGG-16 FT) performed equally well on photos and drawings but still performed poorly on sketches.