Image Matting without Matting-Specific Annotations via Eikonal Fields

Image matting aims to estimate a continuous alpha matte that separates foreground objects from the background. However, existing deep matting methods still rely heavily on task-specific supervision for matting, including alpha mattes and trimaps, which are costly to annotate and limit their applicability in annotation-scarce scenarios. To address this limitation, we propose FreeMatte, a novel annotation-free framework that learns to predict high-quality alpha mattes without requiring any matting-specific annotations. The proposed framework formulates the construction of matting supervision as a foreground-driven wavefront propagation process. Starting from a coarse localization prior generated by an off-the-shelf segmentation foundation model, it solves an Eikonal equation constructed from the prior and the input image itself to obtain an arrival-time field that encodes how readily each pixel can be reached from the coarse foreground source. The arrival-time field is then translated into confidence-aware sparse supervision, which constrains the matting model to learn alpha prediction from the propagation structure encoded by the field. Extensive experiments across diverse benchmarks show that FreeMatte achieves performance competitive with task-supervised matting baselines under this challenging setting, ranking within the top two on 13 out of 20 evaluated dataset-metric pairs.