How the regional coordinates align with the images in the region？
wang-hui111 opened this issue · comments
I would like to ask how NExT-Chat realizes the alignment of region coordinates with images in the region. The information I obtained from the paper is to directly encode coordinates such as x1, y1, x2, y2 and then send them together to llm. However, how does the model perceive which region it is? When training a corpus with coordinates, how the data is processed is something I don't understand.