Question about performance on other datasets

[Jiayi Weng]

I did some experiments. Using Alexnet imagenet pretrained model in pytorch as baseline (delete classifier[6] and add a linear layer).

I use caltech256 first 50 class for the training and testing. For training, I use first 10 images per class. However, the baseline CNN top1 acc is ~67.5%, for the nearest neighbour method, the top1 acc is ~64.7%, and for the parameter prediction approach, the top1 acc is ~70.0%. It seems strange that the baseline is so high, and the parameter prediction's approach seems not strongly outperforms. Could you please give me some advice?