Research map / 3D vision / Depth and stereo

$\alpha$Depth: Learning Single-Pass Soft Boundary Decomposition for Stereo Conversion

NeurIPS 2026 · 3D vision › Depth and stereo · approach: Dense depth prediction and completion

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes αDepth, a single-pass layered representation (alpha, FG/BG colors and depths at soft boundaries) with a Circular Alpha Representation, to improve stereo conversion at hair and defocus-blur boundaries.

Problem

Stereo conversion from monocular images fails at soft boundaries (hair, defocus blur) because single-layer depth and mixed FG/BG colors cause ambiguity, and existing matting methods need user guidance or fail on multi-target scenes.

Approach

Estimate local two-layer (FG/BG) color, depth and alpha only at soft boundaries, and encode alpha on a circle via sin/cos so that 0 and 1 map to the same point, removing the alpha-valley discontinuity at intersecting boundaries; this enables automatic scene-level decomposition in one forward pass.

Key results

• On Marvel-10K stereo video conversion, αDepth reaches S-PSNR 28.46 and FVD 1.72, versus 27.10 and 2.14 for the prior best HairGuard.
• On Mono2Stereo stereo image conversion, S-PSNR 25.60 and LPIPS 0.1181 versus HairGuard's 25.05 and 0.1212.
• On AIM-500 and P3M-10K alpha matting, CAR (with depth guidance, no trimap or mask) achieves average rank 2.17, comparable to ViTMatte (trimap) and behind MatAnyone 2 (mask) at 1.67; the ablation shows CAR beats vanilla alpha on P3M-10K SAD (4.09 vs 4.47).

Surprising findings

• Circular alpha encoding gives comparable matting to trimap- and mask-guided methods without any user guidance, though it does not beat MatAnyone 2 on average rank.
• Matting-based stereo conversion gains on full-image PSNR are small (e.g., 33.31 vs 33.09 on Mono2Stereo), with most gains concentrated on soft regions.

Limitations

• Global scene structure and scale depend on the quality of the input depth map.
• Two-layer representation may not capture three or more overlapping semi-transparent layers.
• Image-based model without explicit temporal constraints may flicker in dynamic video.

Experimental setup

Datasets: AM-2K Distinctions-646 Composition-1K RealEstate10K DL3DV-10K Mono2Stereo Marvel-10K AIM-500 P3M-10K

Benchmarks: Mono2Stereo Marvel-10K AIM-500 P3M-10K

Metrics: S-PSNR S-SSIM LPIPS DISTS FVD SAD Grad Conn PSNR SSIM

Baselines: StereoDiffusion Mono2Stereo StereoCrafter SplatDiff HairGuard ViTMatte GVM MatAnyone 2

Builds on: HairGuard ViTMatte Depth Anything V2 DINOv2 softmax splatting

Closest papers by meaning

• PatchFusion: An End-to-End Tile-Based Framework for High-Resolution Monocular Metric Depth Estimation (CVPR 2024 · similarity 0.86)
• DRIP: Unleashing Diffusion Priors for Joint Foreground and Alpha Prediction in Image Matting (NeurIPS 2024 · similarity 0.85)
• Self-Distilled Depth Refinement with Noisy Poisson Fusion (NeurIPS 2024 · similarity 0.85)
• Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.85)
• Dual-Level Precision Edges Guided Multi-View Stereo with Accurate Planarization (AAAI 2025 · similarity 0.85)
• Depth Pro: Sharp Monocular Metric Depth in Less Than a Second (ICLR 2025 · similarity 0.85)