Rethinking interactive image matting as incremental Gaussian process regression problems

Interactive Image Matting (IIM) aims to predict alpha mattes through user interaction. Traditional methods often depend on user experience to interact at the regions where the alpha matte are inaccurate. However, regions with inaccurate model predictions do not necessarily correspond to areas of high model uncertainty, so these methods are unable to effectively reduce model uncertainty, resulting in low interaction efficiency. To address this issue, we observe a commonality between IIM tasks and Gaussian Process (GP) regression: the former predicts alpha values of unlabeled pixels based on user-labeled information, while the latter predicts observations of unknown data based on known data and provides uncertainty estimation for predictions. Based on this observation, we model IIM as an incremental GP regression problem and propose a novel IIM paradigm, IIM-GP. First, IIM-GP is the first model to incrementally utilize model-predicted uncertainty to guide user interaction and update matting results, significantly enhancing interaction efficiency and prediction reliability. Second, an incremental update strategy is implemented within the GP framework, overcoming traditional GP models’ inefficiency in updating results for IIM tasks. Additionally, IIM-GP employs a strategy of selecting $p$ inducing points from $n$ labeled pixels to perform variational inference on GP, reducing computational complexity from $O\left(n^3\right)$ to $O\left(n{p}^2\right)$ ($p\ll n$). Comprehensive experiments on five widely-used datasets (Composition-1k, AIM-500, Distinctions-646, HIM2K and AM-2K) demonstrate that IIM-GP achieves competitive performance.