Abstract
The Dolly Zoom, or "Vertigo effect," is a classic cinematic technique that creates a dramatic perspective shift by simultaneously moving the camera and adjusting the focal length. Traditionally requiring specialized equipment, this effect is difficult to achieve on mobile devices. This paper introduces a computational photography pipeline that synthesizes the Dolly Zoom from a single static image and its corresponding depth map. By employing camera geometry modeling and depth-aware inpainting, the proposed method generates high-quality cinematic sequences suitable for smartphone hardware.
Introduction
The Dolly Zoom keeps a foreground subject at a constant size while the background appears to shrink or expand. This creates a powerful sense of unease or realization. On smartphones, the lack of mechanical precision in lens tracking makes physical execution nearly impossible. While existing software solutions often require video input or specialized hardware, this research focuses on generating the effect from a single static shot.
Methodology
The synthesis pipeline is built on 3D computer vision principles and camera geometry.
Camera Geometry Modeling
The system uses a pinhole camera model to derive closed-form equations that map pixels from an initial capture to a synthesized "virtual" camera position. This allows for the mathematical simulation of camera movement along its principal axis.
Single vs. Dual Camera Input
For single-camera setups, digital zoom and image warping simulate movement. On dual-camera devices, the pipeline leverages the wider field-of-view (FoV) from a secondary lens to fill in occluded background areas that become visible during virtual camera translation.
Depth-Aware Inpainting
Warping based on depth creates "holes" or occlusions. The authors developed a two-stage algorithm to address this:
- Depth Map Completion: Fills holes in the depth map using a nearest-neighbor approach prioritizing background values.
- Texture Propagation: Fills image holes by propagating textures from the background toward the foreground, preventing "bleeding" of foreground colors into the background.
Shallow Depth of Field (SDoF)
A synthetic blur is applied as a final step. This enhances the cinematic aesthetic and serves a functional purpose by masking minor artifacts from the warping and inpainting process.
Experimental Results
The method was evaluated using both a synthetic Unity 3D dataset and real-world smartphone captures.
Quantitative Performance
Using the synthetic dataset, the authors measured Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Although metrics decrease with extreme zoom levels due to increased inpainting requirements, the application of SDoF improved scores by smoothing reconstruction noise.
Qualitative Success
Qualitative results demonstrate that foreground subjects remain sharp and stationary while the background undergoes the characteristic Dolly Zoom "stretch."
The pipeline is optimized for mobile efficiency, favoring geometric warping over heavy deep-learning models.
Conclusion
The effect is constrained by the physical Field of View of the hardware; background data not captured by any lens cannot be synthesized. Additionally, complex occlusions like tree branches remain challenging. Future work may involve integrating AI-driven occlusion handling to further improve image quality.
Summary updated