Traditional multi-camera systems rely on a simple geometric requirement: overlap. To determine the spatial relationship between two cameras, they must typically "see" the same thing at the same time. By identifying shared feature points, researchers can use epipolar geometry to triangulate positions and orientations.
However, many modern robotic and surveillance configurations—such as 360-degree autonomous vehicle rigs or distributed indoor camera networks—are designed specifically to minimize overlap to maximize coverage. When cameras point in opposite directions, the standard geometric toolbox fails. To solve this, researchers have developed methods to bridge these "blind" gaps by leveraging motion, environmental anchors, and optical redirects.
Motion as a Geometric Bridge
The most robust way to calibrate non-overlapping cameras is to move them. If two cameras are rigidly attached to the same platform, their individual ego-motion paths are not independent; they are constrained by the fixed transformation (the extrinsics) between them.
As Junzhou Chen et al. demonstrate, even without shared visual features, the relationship between a camera's rotation and translation can be used to isolate the fixed offset of a joint or mounting point. This is often framed as a "hand-eye" calibration problem. By observing how different sensors move through space relative to one another, the system can calculate spatial offsets without ever needing a shared visual target.
This principle extends to LiDAR systems as well. Atsushi Suzuki et al. show that by leveraging robot motion to "stitch" together observations of the floor and surrounding objects over time, a mobile robot can achieve sub-centimeter calibration accuracy across sensors that never share a field of view. The intuition is that the environment itself becomes a persistent reference frame that the sensors visit at different moments.
Anchoring to Global Infrastructure
If the cameras cannot see each other, they can instead be localized against a shared external reference. This "infrastructure-based" approach treats the world as a giant calibration target.
One method involves using a pre-existing 3D map. Marc Pollefeys et al. and Marcel Geppert et al. propose matching image features from individual cameras to a known map of the environment. Even if Camera A and Camera B see entirely different walls, their shared registration to the same global map implicitly defines their relative pose.
In controlled settings, researchers like Chuan Fang et al. use "calibration rooms" covered in markers. Because the markers' positions are known globally, a single shot from a 360-degree rig is enough to calibrate every sensor, as each camera sees a different but known subset of the room's patterns.
Mobile Agents as Temporal Links
In distributed networks, such as cameras mounted on different walls of a building, cameras may be separated by literal obstacles. In these cases, a "mobile agent"—a person or a robot—can act as a bridge.
Sosuke Yamao et al. describe a system where a mobile agent's trajectory serves as a "mobile keypoint." As the agent walks through the building, different cameras observe it at different times. By aligning these local observations with the agent's globally optimized trajectory, the system can register isolated cameras into a unified spatial network. Similarly, Jianping Yin et al. use humans moving through a scene as dynamic calibration objects, leveraging human pose estimation to provide the geometric constraints that raw pixels cannot.
Virtual Overlap via Optical Redirects
A more direct, hardware-based solution involves using mirrors to "bend" the field of view. By placing a planar mirror in front of a camera, a researcher can reflect a calibration pattern—or even another camera—into the sensor's view.
Ram Krishan Kumar et al. showed that a simple mirror can derive linear constraints to recover the 3D position of cameras in a single coordinate frame. Andrey Bushnevskiy et al. took this further by using mirrors to make cameras "see" each other's epipoles. This creates a virtual overlap, allowing standard geometric solvers to work on cameras that are physically pointing in opposite directions.
Learning Hidden Geometric Cues
Recent advances in deep learning suggest that even without explicit motion or mirrors, there may be enough information in a single image to infer its relationship to another disjoint view. Ruojin Cai et al. demonstrated that neural networks can learn to exploit "hidden" cues like vanishing points, shadow directions, and scene symmetries to predict the relative rotation between two images with zero overlap. This suggests that the statistical structure of our world—such as the fact that gravity usually points "down" and walls are usually vertical—provides a latent coordinate system that can be used for alignment.
Key Foundational Work
Calibration of a Hybrid Camera Network (2003) — contributed 1 concept elements to this direction.
Alignment of Non-Overlapping Sequences (2001) — contributed 1 concept elements to this direction.