Torchvision pretrained models perform bad on very simple images classification

[Fnjn (Fanjin Zeng)]

I tried torchvision pretrained models vgg16_bn and resnet50 to classify simple images, but they perform really bad.

Here are their performance examples:
vgg16_bn

resnet50

I cant do transfer learning, because I dont have any label for these images. Anyone know how to improve the performance? Thanks

[bobey (mubarek mohammed)]

Normalize the images…probably do some more valid transformations like …standard deviation