Closed GanjinZero closed 1 year ago
Hello,
Actually yes, the best performing model do not encode the questions. MVQA task right now is a bit of a dummy task. Questions are like "what is the abnormality in the image" most of the time. There are some recent paper that try to create better dataset: https://openreview.net/pdf?id=uH_RlkvQMUs
It seems MVQA inputs a image and do label classification? Question information is discarded?