Inverse depth parametrization is a method used in computer vision and simultaneous localization and mapping (SLAM) to represent the position of a 3D point in a camera's view. Instead of storing the point's Euclidean coordinates (x, y, z) directly, it stores the point's direction (azimuth and elevation angles) and the inverse of its depth (1/z). This representation is particularly valuable for monocular SLAM systems, where depth is not directly observable and must be estimated from motion over time.
The core advantage of inverse depth parametrization lies in its treatment of uncertainty. For distant points, a small error in depth corresponds to a large error in the Euclidean z-coordinate, making Gaussian noise models in standard Cartesian coordinates poorly suited. In inverse depth, the uncertainty distribution becomes more symmetric and closer to Gaussian, which aligns well with the Kalman filter and other estimation frameworks commonly used in SLAM. This leads to more stable and accurate state estimation, especially during the initial frames when a feature's depth is highly uncertain.
Historical Context
The technique gained prominence in the mid-2000s through the work of Javier Civera, Andrew Davison, and José María Martínez Montiel, who published influential papers on monocular SLAM with inverse depth parametrization. Their 2008 paper "Inverse Depth Parametrization for Monocular SLAM" (published in IEEE Transactions on Robotics) formalized the approach and demonstrated its effectiveness in real-time tracking. This work built on earlier research in Computer vision and Machine learning that explored alternative parametrizations for 3D reconstruction, but inverse depth proved particularly well-suited for the delayed and undelayed initialization strategies used in monocular systems.
Prior to this, monocular SLAM systems often struggled with features that were far from the camera or that had large depth uncertainty. The inverse depth formulation allowed these systems to initialize features with a very large depth variance (effectively modeling infinite depth) and then refine the estimate as the camera moved and provided parallax. This was a significant improvement over earlier methods that required features to be initialized with a fixed depth or that discarded features with high uncertainty.
Mathematical Formulation
In inverse depth parametrization, a 3D point is represented as a six-dimensional state vector:
- The camera position at the first observation (x0, y0, z0)
- The azimuth angle (θ) and elevation angle (φ) defining the ray direction from that position
- The inverse depth (ρ = 1/d)
This representation is a minimal parametrization of the point's location, avoiding the need for a full 3D coordinate with redundant information. The actual 3D coordinates can be recovered by moving from the initial camera position along the ray direction by a distance of 1/ρ.
The choice of inverse depth rather than depth itself is motivated by the observation that for a camera moving with constant velocity, the inverse depth of a static point evolves linearly in the absence of noise. This linearity simplifies the prediction step in filtering algorithms and improves numerical conditioning. Additionally, the inverse depth can be initialized to zero (representing a point at infinity) without causing singularities, which is not possible with a standard depth parametrization.
Applications in SLAM and Visual Odometry
Inverse depth parametrization has become a standard component in many monocular SLAM systems. It is used in both filtering-based approaches, such as the original MonoSLAM system developed at the University of Oxford, and in keyframe-based optimization methods. For example, the PTAM (Parallel Tracking and Mapping) system, developed by Georg Klein and David Murray, adopted inverse depth for its mapping thread, and subsequent systems like LSD-SLAM and ORB-SLAM have incorporated similar ideas in their initialization and mapping stages.
The technique is particularly useful in scenarios where the camera undergoes significant translational motion, providing the parallax needed to reduce depth uncertainty. In visual odometry for autonomous vehicles, such as those developed by Waymo or Tesla, inverse depth helps maintain consistent 3D maps from monocular camera streams, even when features are at varying distances. It also plays a role in augmented reality and structure-from-motion pipelines, where accurate depth estimation from a single moving camera is essential.
Advantages and Limitations
A key advantage of inverse depth parametrization is its ability to handle features at very large distances, including points that are effectively at infinity. In standard Euclidean parametrization, such points cause the covariance matrix to become ill-conditioned, leading to numerical instability. Inverse depth avoids this by keeping the parameter bounded and the uncertainty well-modeled.
Another benefit is the ease of initialization. A new feature can be initialized with a large inverse depth variance, representing complete ignorance about its distance, and then updated as more observations are made. This contrasts with methods that require an initial depth estimate or that use delayed initialization to gather enough parallax first.
However, the parametrization is not without drawbacks. The representation requires storing the initial camera position for each feature, which increases memory usage. It also introduces a coupling between the camera pose and the feature parameters, which can complicate the optimization process. In practice, many systems switch to a Euclidean parametrization once a feature's depth uncertainty has been sufficiently reduced, to simplify the state representation and reduce computational cost.
Relation to Other Techniques
Inverse depth parametrization is closely related to other techniques in probabilistic robotics and computer vision. It shares conceptual similarities with the use of disparity in stereo vision, where the inverse of depth is directly proportional to the disparity between left and right camera images. In Deep learning approaches to monocular depth estimation, such as those using Neural network architectures, the output is often represented as inverse depth or log-depth to improve training stability and accuracy, reflecting the same underlying statistical reasoning.
The technique also connects to broader estimation frameworks like the Kalman Filter and particle-filter, which are used in SLAM and tracking. The choice of parametrization is a critical design decision in these systems, and inverse depth has proven to be a robust and practical choice for monocular vision, influencing a wide range of subsequent research and commercial implementations.