whwu95/GPT4Vis

Leverages GPT-4's multimodal capabilities for zero-shot classification by generating rich semantic descriptions for class labels and processing visual inputs (images, videos, point clouds) through GPT-4V's vision API. The approach evaluates performance across 16 benchmark datasets spanning three modalities, providing pre-generated prompts and reproducible inference scripts alongside comprehensive results and ground-truth annotations.

185 stars. No commits in the last 6 months.