The Stereoscopic Zoom

Sergi Pujades, Frederic Devernay, Laurent Boiron, Rémi RonfardWorkshops @ CVPR 2017 · DOI

Introduction

In traditional 2D filmmaking, long focal length lenses are indispensable tools for directors, allowing them to bring distant action closer or create specific aesthetic perspective distortions. However, in stereoscopic (3D) cinema, the use of zoom lenses introduces significant technical challenges. Standard two-camera rigs often force a compromise between depth perception and viewer comfort, leading to either "visual fatigue" or the undesirable "cardboard effect."

The Stereoscopic Challenge

When capturing 3D content with a long lens, filmmakers face a dual constraint. To maintain a realistic sense of volume, a wide baseline (distance between cameras) is often required. However, this can cause ocular divergence, where the viewer's eyes must turn outward to fuse the images, leading to significant eye strain. Conversely, reducing the baseline to ensure comfort often results in the "cardboard effect," where 3D objects appear as flat cutouts at varying depths rather than solid volumes.

Geometry of stereoscopic acquisition and projection.

As illustrated in the geometry above, the relationship between acquisition parameters (like camera baseline $b$ and convergence distance $H$) and projection parameters (like eye distance $b'$ and screen distance $H'$) is critical for maintaining the perceived depth $z'$ relative to the real depth $z$. The authors aim to break this trade-off by providing a way to generate stereoscopic zoom shots that are both comfortable to watch and rich in perceived volume.

Proposed Multi-Camera Solutions

To overcome these limitations, the authors propose moving beyond the traditional binocular rig toward multi-camera configurations combined with Image-Based Rendering (IBR).

The Quadri-Rig: Virtual Proximity

The Quadri-Rig model aims to simulate virtual camera positions that would be physically impossible or impractical to achieve. By using four physical cameras placed at a distance, the system captures both the subject of interest and the background from multiple angles. These views are then blended to synthesize two virtual views that mimic a close-up shot with a natural, comfortable baseline.

The Tri-Rig: 3D World Distortion

The Tri-Rig offers a more computationally innovative approach by using three cameras: one "master" camera for the left eye and two auxiliary cameras to provide depth information.

Multi-camera configuration for subject and background capture.

Instead of simply synthesizing new camera positions, the Tri-Rig employs a world-distortion function ($\Phi$). This function distorts the 3D coordinates of the scene before they are projected onto the screen. By implementing this distortion within a vertex shader, the system can handle complex occlusions and depth testing efficiently on standard GPU hardware. This approach is particularly clever because it keeps the left eye's image as a "raw" capture from a real camera, resulting in a higher perceived final quality.

Key Concepts: The Roundness Factor

A central metric in this research is the "Roundness Factor" ($\rho$). The authors define this mathematically to quantify the perceived volume of an object. When $\rho$ falls below a threshold (approximately 0.3), the human brain begins to perceive objects as flat. This metric allows for the formal calculation of the minimum baseline required to maintain "roundness" without exceeding the limits of visual comfort.

Results and Performance

The methods were evaluated using the "blender lego" synthetic dataset, providing a ground truth for camera positions and scene geometry.

  • Accuracy: The Quadri-Rig achieved high fidelity with a Peak Signal-to-Noise Ratio (PSNR) of approximately 34.0 dB.
  • Occlusion Handling: The Tri-Rig proved superior in handling occlusions, significantly reducing "black holes" in the rendered images where no camera data was available.
  • Efficiency: The use of vertex shaders for world distortion suggests that these methods are viable for near-real-time applications in live broadcast or interactive media.

Conclusion and Future Work

The "Stereoscopic Zoom" provides a robust framework for high-quality 3D zooming. While the current models assume a simplified "subject + background" layout, future iterations could expand to more complex scenes. Additionally, while the mathematical foundations are sound, further subjective user studies will be essential to validate the aesthetic preference for these distorted 3D views in professional cinematography.

Summary updated

BibTeX
@inproceedings{pujades2017stereoscopic,
  author = {Sergi Pujades and Frederic Devernay and Laurent Boiron and Rémi Ronfard},
  title = {{The Stereoscopic Zoom}},
  booktitle = {Workshops @ CVPR},
  year = {2017},
  doi = {10.1109/cvprw.2017.170},
}
Figure