Spot Anything: Subject‐Agnostic Background Matting

Background matting, which incorporates an auxiliary background image, provides an efficient and accurate solution for alpha matte estimation in natural scenes. While designed to extract arbitrary subjects, prior works have concentrated on specific categories, such as humans or salient objects. Moreover, reliance on matte‐based image synthesis restricts the variety of available training data. This paper presents a subject‐agnostic background matting framework that precisely localizes subjects of any category by exploiting foreground–background contrast in the feature domain. We propose a diff detection module (DDM) to enhance accuracy for fine‐grained and small‐scale targets through differential feature analysis. To reduce the category bias between limited training sets and the broader real world, we design a self‐supervised strategy that leverages the segment anything model (SAM) for segmentation and employs inpainting to generate pseudo‐mattes. In addition, we release SynBM‐10K, the largest synthetic general‐purpose background matting dataset to date, and RWCS‐100, a manually annotated test set for real‐world complex scenarios. Experimental results show that our method markedly improves generalization over existing approaches, extending the applicability of background matting to arbitrary objects while maintaining high processing speed and accuracy.