Export
Research Overview
Geoffrey Cross focuses on multiple view geometry, 3D surface reconstruction, and stereo video segmentation. His early research established foundational methods for recovering quadric and smooth surfaces from perspective views using epipolar and trifocal geometry. His later collaborative work addresses real-time bilayer segmentation by fusing stereo matching with color and contrast models.
Research Trajectory
Geoffrey Cross’s documented research trajectory begins in the late 1990s with a primary focus on the mathematical foundations of 3D computer vision, specifically the reconstruction of surfaces from multiple perspective views. As a lead author during this period, his work addressed the geometric ambiguities inherent in inferring three-dimensional shapes from two-dimensional images. In Quadric Surface Reconstruction from Dual-Space Geometry, Cross demonstrated how a one-parameter family of quadrics could be recovered from outlines in two views, and how point correspondences or degenerate quadrics could systematically reduce reconstruction ambiguity. This work provided theoretical grounding for applications in 3D modeling and image registration.
Continuing this focus on multi-view geometry, Cross led the development of new techniques for analyzing smooth surfaces. In Parallax Geometry of Smooth Surfaces in Multiple Views., he investigated scenes containing smooth surfaces and a reference plane. This research introduced novel solutions for computing epipolar and trifocal geometry, notably demonstrating that the epipole can be determined from bitangents between homography-registered occluding contours. The paper also provided algorithms for automatically estimating fundamental matrices, trifocal tensors, and camera matrices from these specific scene types.
In the mid-2000s, Cross's published work shifted toward collaborative research in video processing and real-time computer vision. Operating as a contributing middle author, he participated in the development of robust segmentation algorithms for binocular stereo video. This work is detailed in Bi-Layer Segmentation of Binocular Stereo Video and its subsequent journal expansion, Probabilistic Fusion of Stereo with Color and Contrast for Bilayer Segmentation.. These papers addressed the known error rates of relying solely on color, contrast, or stereo matching alone for foreground-background separation. The research team proposed fusing these modalities to infer layers accurately in real time. A key methodological contribution of this collaborative effort was layered dynamic programming (LDP), which solved stereo matching in an extended six-state space representing foreground layers, background layers, and occluded regions.
Current Directions
The most recent papers in the provided evidence are from 2005 and 2006, centering on probabilistic fusion techniques for stereo video segmentation. This work investigates how combining multiple visual cues—specifically color, contrast, and stereo disparities—can overcome the limitations of using any single modality for foreground-background separation. The integration of these cues via layered dynamic programming and on-the-fly color modeling represents the latest documented focus of his research trajectory, emphasizing real-time applicability and robust probabilistic modeling in dynamic video environments.
Profile updated
Loading new papers…
Loading research neighbors…
See something wrong? Report a mistake