1 paper · 1 filter
Guanhua Chen, Yutong Yao, Shenghe Sun +5
Recent advances in vision-language models (VLMs) have achieved impressive results on standard image-text tasks, yet their potential for visual procedure question answering (VP-QA)…