Datasets:
Clarification on the Absence of Question Type Annotations in PathVQA
Thank you for releasing the PathVQA dataset. We have been using PathVQA as an evaluation benchmark in our recent work on medical visual question answering, and we greatly appreciate the effort put into building and maintaining this dataset.
While reviewing the released version on Hugging Face (flaviagiammarino/path-vqa) as well as the data processing script, we noticed that the dataset does not explicitly provide question type annotations (e.g., open-ended vs. closed-ended / yes-no). However, in the paper, questions are described as being categorized into open-ended and closed-ended types.
We would like to kindly ask for clarification on the following points:
Are the question type annotations available in any version of the dataset that is not publicly released?
If not, was the categorization performed only at a statistical or conceptual level in the paper rather than provided as explicit labels?
For evaluation purposes, would you recommend inferring question types heuristically (e.g., based on answer values such as “yes/no”), or treating all questions as open-ended?
This clarification would be very helpful for ensuring a fair and faithful evaluation protocol when comparing with previously reported results.
Thank you again for your valuable contribution to the pathology VQA community. We look forward to your guidance.