Reproduce “Infinite-Homography as Robust Conditioning for Camera-Controlled Video Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Min-Jung Kim, Jeongho Kim, Hoiyeong Jin, Junha Hyung, Jaegul Choo
InfCam presents a robust, depth-free framework for camera-controlled video generation that leverages infinite homography warping and a comprehensive data augmentation pipeline to synthesize dynamic scenes. The system achieved a Rotation Error of 3.162 and Translation Error of 0.438 on the WebVid dataset, demonstrating superior camera-pose fidelity and visual quality compared to prior methods.
Key insights
- Explicitly encoding 3D camera rotation via depth-free infinite homography warping provides a robust, noise-free conditioning signal that significantly improves pose fidelity by constraining the search space for parallax.
- Decomposing the reprojection task into a known rotational warp and a learned residual parallax enables the generative model to implicitly refine 3D geometry and depth through end-to-end training.
- Targeted data augmentation, specifically varying initial camera poses and focal lengths, is crucial for mitigating dataset biases and achieving strong generalization across diverse camera trajectories and fields of view.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.