A key question for understanding multimodal performance is analyzing the ability for a model to have basic
vs. detailed understanding of images. These capabilities are needed for models to be used in
real-world tasks, such as an assistant in the physical world. While there are many dataset for object detection
and recognition, there are few that test spatial reasoning and other more targeted task such as visual prompting.
The datasets that do exist are static and publicly available, thus… See the full description on the dataset page:
https://huggingface.co/datasets/microsoft/IMAGE_UNDERSTANDING.
