Skip to main content

Table 9 Other architectures

From: Pre-Training Without Natural Images

Arch. PT PTimg Type C10 C100 IN1k P365 VOC12 OG
AlexNet Scratch 88.2 64.5 56.5 49.7 61.5 15.7
P365 Nat. Img Supervision 91.9 72.5 46.3 N/A 22.0
IN1k Nat. Img Supervision 93.6 76.0 50.5 77.9 23.3
F1k Formula Formula-supervision 89.4 65.91 59.0 50.5 65.5 25.2
F10k Formula Formula-supervision N/A N/A 57.6 49.6 52.8 21.5
ResNet-18 Scratch 89.3 67.5 69.7 51.6 64.2 0.5
P365 Nat. Img Supervision 93.2 73.4 65.5 76.7 14.7
IN1k Nat. Img Supervision 95.7 81.6 50.2 83.4 20.9
F1k Formula Formula-supervision 92.1 72.4 65.2 50.7 64.2 33.0
F10k Formula Formula-supervision 91.9 72.7 64.8 51.7 70.7 30.9
ResNet-50 Scratch 87.6 62.7 76.1 49.9 58.9 1.1
P365 Nat. Img Supervision 94.2 76.9 71.4 78.6 10.5
IN1k Nat. Img Supervision 96.8 84.6 50.3 85.8 17.5
F1k Formula Formula-supervision 93.4 75.7 70.3 49.5 58.9 20.9
F10k Formula Formula-supervision 94.1 77.3 71.5 50.8 73.6 29.2
ResNet-152 Scratch 85.2 59.5 78.3 48.2 55.1 1.6
P365 Nat. Img Supervision 94.9 77.8 71.4 55.4 15.8
IN1k Nat. Img Supervision 97.7 86.8 49.6 87.4 20.1
F1k Formula Formula-supervision 93.7 75.7 71.8 54.3 55.1 23.4
F10k Formula Formula-supervision 94.3 78.0 72.1 54.4 73.2 30.6
ResNeXt-101 Scratch 85.3 59.7 68.7 46.9 55.9 2.4
P365 Nat. Img Supervision 94.8 78.3 72.3 55.0 22.6
IN1k Nat. Img Supervision 96.0 81.2 48.1 81.6 25.5
F1k Formula Formula-supervision 94.0 77.1 72.7 48.9 55.9 24.0
F10k Formula Formula-supervision 94.3 78.3 72.3 48.6 73.4 27.6
DenseNet-161 Scratch 90.9 69.4 77.1 51.0 63.8 2.7
P365 Nat. Img Supervision 95.3 78.8 73.6 79.3 25.6
IN1k Nat. Img Supervision 97.6 86.7 51.8 86.4 28.5
F1k Formula Formula-supervision 94.0 77.0 72.3 51.1 74.1 23.5
F10k Formula Formula-supervision 94.5 78.8 72.6 52.0 73.1 28.9
  1. We list AlexNet, ResNet-{18, 50, 152}, ResNeXt-101, and DenseNet-161 architectures. We compare the FractalDB-1k/10k (F1k/F10k) with training from scratch (Scratch), Places-365/ImageNet-1k pre-trained models (P365 / IN1k). The datasets used are as per Table 8. The bold and underlined values show the best scores, and bold values indicate the second best scores