How much impovement does VG dataset bring?

MILVLG / mcan-vqa

Deep Modular Co-Attention Networks for Visual Question Answering

Apache License 2.0

438 stars 88 forks source link

Closed eustcPL closed 5 years ago

MIL-VLG commented 5 years ago

Without VG, the small MCAN model report an overall accuracy of 70.26%, which is about 0.4% lower than that with VG.