More datasets ideas

[arthurdouillard]

• [x] Food101, classif, implem in #191
• [X] CIFAR10
• [X] CIFAR100
• [X] Birdsnap, must improve the crawler with multi-threading and python3 , implem in #219
• [X] SUN397, implem in #223
• [x] Stanford Cars, implem in #198
• [X] FGVC Aircraft, implem in #201
• [x] Pascal VOC 2007 classif, implem in #202
• [X] Describable Textures DTD
• [x] Oxford pets
• [x] Caltech 101, Caltech 256, implem in #189
• [x] Oxford flowers 102
• [X] MNIST
• [X] Facial emotion reco 2013, implem in #221
• [x] STL10, implem in #186
• [X] EuroSAT, implem in #220
• [X] GTSRB, implem in #224
• [X] ImageNet

Datasets that will be difficult to implement, because no scenarios can handle them yet or not available. They won't be implemented anytime soon.

• PatchCamelyon, format is in big separate h5 files, but not the format of h5 we used, so either we would need to reformat the dataset (more data generated stored on disk) or change a lot the code (but not worth it for a single dataset)
• RESISC45, seems to use a Microsoft Windows archive format, not really worth it to make it work on linux
• Country211, seems internal to OpenAI and super bad
• UCF101, video #204
• Kinetics700, video #204
• CLEVR Counts, multimodal #203!
• Hateful Memes, multimodal #203!
• Rendered SST2, multimodal? #203, seems internal to OpenAI
• KITTI, need to have support of object detection. Btw how to handle scenario? Like in segmentation (overlap/disjoint)? See #194

[arthurdouillard]

Md Rifat Arefin on Oxford flowers & Oxfords pets

[arthurdouillard]

I'm taking Birdsnap

[arthurdouillard]

I'm taking eurosat and sun397