search

uqzhichen / SDGZSL

[ICCV2021] Official Pytorch implementation for SDGZSL (Semantics Disentangling for Generalized Zero-Shot Learning)
36 stars 8 forks source link

Dataset #1

Closed KernLC closed 2 years ago

KernLC commented 2 years ago

"missing the CUB dataset." How can I get the visual features of CUB dataset(not the visual features extracted from the pre-trained deep models)

uqzhichen commented 2 years ago

Hi Kern,

The data.mat file is the visual features for CUB dataset. You can refer to the dataloader how to load the visual features.

Regards, Zhi

KernLC commented 2 years ago

Hi Kern,

The data.mat file is the visual features for CUB dataset. You can refer to the dataloader how to load the visual features.

Regards, Zhi

Thanks for your reply! I have downloaded the dataset. But I was confused by the "data.mat", could you please explain the meaning of each filed in the "data.mat", such as "seen_pro", "train_fea", "train_pro" etc.

uqzhichen commented 2 years ago

The dataset is actually from EPGN(CVPR2020) https://github.com/yunlongyu/EPGN, which is mentioned in the README file. seen_pro is the class embeddings for seen classes, train_fea is the visual features of seen classes. You can refer to the dataloader for more details. https://github.com/uqzhichen/SDGZSL/blob/beed9df03d6f7f0ce6e71753b346da670345a861/dataset_GBU.py#L97-L108