Unlocking Self-Supervision for Stereoscopic Video Generation

ICML 2026

Jingyi Lu, Kai Han
Visual AI Lab, The University of Hong Kong

Code

Demo

Gallery

BibTeX

	Scalability	Real-World	Quality
Real Stereo Pairs	✗	✓	✗
Synthetic Data	✓	✗	✓
Geometric Reciprocity (Ours)	✓	✓	✓

Geometric Reciprocity constructs scalable, high-quality stereo inpainting data from real-world monocular videos, using inference-time consistent masks without relying on paired stereo videos. ~300K Kinetics-GRT Masks

Method Insight

Depth-image-based rendering (DIBR) uses monocular depth estimates to warp the left view into the right view, but newly exposed regions lack ground-truth supervision.

Starting from cycle consistency, GRT shows that pixels failing to complete the target-to-source-to-target round trip exactly reveal the target-view disocclusions, yielding supervision masks for stereo inpainting.

Dataset Gallery

Kinetics videos paired with GRT-derived masks form training samples for stereo video inpainting. Compute masks for your own videos

Results Gallery

Results after training a stereo video inpainting model with GRT-constructed data.

BibTeX

@inproceedings{lu2026grt,
  author    = {Jingyi Lu and Kai Han},
  title     = {Geometric Reciprocity: Unlocking Self-Supervision for Stereoscopic Video Generation},
  booktitle = {International Conference on Machine Learning (ICML)},
  year      = {2026},
}