Segment Anything Models (SAMs) like SEEM and SAM have achieved great performance on various downstream datasets at the cost of crafting spatial and semantic prompts. Previous prompt learning methods can learn prompts automatically but largely focus on learning semantic prompts, while how to learn effective spatial prompts that are important to SAMs is largely under-explored. Inspired by Hough Voting that detects a complex object by voting from its parts, we propose Hough Voting-based Spatial Prompt Learning (HoughSpaPL) that designs three types of voting mechanisms to learn three distinct spatial prompts for different subregions of the visual concept (e.g., things and stuff), which capture complementary spatial clues and vote together to guide SAMs to generate a precise segmentation mask for the visual concept. Following the same philosophy, we design Hough Voting-based Semantic Prompt Learning (HoughSemPL) that learns distinct semantic prompts for different sub-regions of the visual concept, which capture complementary semantic clues and vote together to predict a accurate semantic label for the generated mask. Extensive experiments show that our proposed techniques achieve superior prompt learning performance over popular segmentation datasets. Codes will be released.
