yewsiang/ConceptBottleneck

Traveling Birds Dataset

Closed this issue · 5 comments

Hello,

did you use all images in the train and test folders for training and testing with the Traveling Birds dataset or did you only use those assigned to train/test by CUB200?

Greetings

Hmm, I tried the dataset with a vanilla resnet50 and end up around your numbers when I respect the CUB200 split and achieve >60% accuracy when using the full train and testset provided in the folders.

How certain are you about your split?

And off-topic: Are you aware of other work that uses this dataset? All the citations of the paper make it hard to filter for the dataset :)

Have you used the dataset processing instructions in https://github.com/yewsiang/ConceptBottleneck/tree/master/CUB? We randomly sample 20% of the train split of CUB200 to use for validation. The CUB_processed folder then contains the metadata for all train/ val/ test splits

For Traveling Birds dataset, we follow the same split as CUB200 (i.e. validation data is included). See this script for more details https://github.com/yewsiang/ConceptBottleneck/blob/master/CUB/scripts/experiments.sh#L95

Okay, I see.
That makes sense, thank you