Title: {\alpha}Depth: Learning Single-Pass Soft Boundary Decomposition for Stereo Conversion

Abstract:

Stereo conversion faces a significant hurdle in accurately modeling soft boundaries, such as hair strands and defocus blur, because the ambiguous interplay between foreground and background complicates the process. Traditional depth models typically generate a single-layer depth map, which creates uncertainty regarding depth correspondence at these soft edges. Although matting techniques can derive opacity for layered structures, they often falter in complex environments containing multiple subjects and generally depend on manual user input. To address this, we present {\alpha}Depth, a layered representation framework designed to decompose soft boundaries for high-fidelity stereo conversion. Our approach first tackles the joint ambiguity of color and depth by estimating distinct layered values at soft boundaries. For scenes with multiple complex targets, we introduce the Circular Alpha Representation (CAR), which transitions the focus from global target extraction to local boundary decomposition. In contrast to earlier matting methods limited to binary foreground/background distinctions, CAR facilitates efficient, scene-level inference without the need for manual guidance. Comprehensive evaluations confirm that {\alpha}Depth delivers state-of-the-art results in stereo conversion, effectively removing background bleeding and structural distortions at soft boundaries.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC