Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation

TL;DR

A self-supervised framework for monocular-to-stereo conversion, where the key is the Geometric Reciprocity Theorem (GRT) that makes disocclusion mask computation consistent between training and testing.

Problem

The critical bottleneck in monocular-to-stereo conversion is disocclusion inpainting. Existing training-based methods rely on scarce stereo pairs or synthetic data with domain gaps.

Approach

The paper derives the Geometric Reciprocity Theorem (GRT), proving that under nearest-neighbor DIBR, the disocclusion mask for synthesizing a target view equals the mask of pixels lost when warping back from target to source. This allows analytical computation of test-time disocclusion masks from monocular images alone, enabling self-supervised learning via cycle consistency.

Results & Contribution

The method achieves substantial improvements over both training-free and supervised state-of-the-art methods. By enabling learning from unlimited monocular videos, it fundamentally addresses the data scarcity problem in stereoscopic video synthesis.