Traditional computer vision requires a handful of point matches—usually five to eight—to determine how a camera has moved between two frames. However, a new direction in geometric vision demonstrates that we can recover camera pose using as few as one or two matches. By replacing simple point coordinates with affine correspondences (which capture local shape) and monocular depth priors (which provide structural context), researchers are creating "minimal solvers" that are orders of magnitude faster and more robust than classical methods.
The Information Density of a Single Match
A standard point match tells you where a feature moved. An affine correspondence (AC) tells you how the area around that feature warped. This local deformation encodes the surface's orientation and the camera's change in perspective.
As Daniel Barath and Levente Hajder demonstrate, this high information density allows for the recovery of motion from a single match if the motion is constrained, such as a vehicle moving on a 2D plane. Banglei Guan and colleagues further show that these single-match solvers can even recover "hidden" variables like the camera's focal length, enabling self-calibration on the fly. In the absolute pose case, Jonathan Ventura and Torsten Sattler show that one AC is sufficient to localize a camera against a 3D map by leveraging the local geometric warping of the feature patch.
Integrating Monocular Depth Priors
While affine correspondences provide local geometry, modern deep learning provides a global geometric prior through monocular depth estimation. Even if these depth maps aren't perfectly "metric" (accurate in meters), they provide a powerful constraint on the relative distance of points.
Ivan Eichhardt introduced the 1AC+D solver, which combines one affine correspondence with deep-learned depth to eliminate degrees of freedom in the pose problem. Similarly, Marc Pollefeys and Petr Hruby have shown that integrating these depth priors with the local surface orientation from an AC allows for accurate pose recovery even in semi-calibrated systems where the camera's internal parameters are unknown.
Why Minimal Solvers Matter: Speed and Robustness
The primary benefit of reducing the number of required matches is the efficiency of RANSAC (Random Sample Consensus). RANSAC works by repeatedly picking a random subset of matches to find a "consensus" pose.
- Traditional 5-point solver: Requires many iterations to find a set of 5 points that are all correct (inliers).
- 1-AC solver: The number of iterations required grows linearly, rather than exponentially, with the outlier ratio.
This shift makes robust estimation nearly instantaneous. Eduardo Arnold and Victor Adrian Prisacariu apply this to "map-free" relocalization, where a camera's pose is found relative to a single reference image without a pre-built 3D map. By lifting 2D matches into 3D using predicted depth, they enable "instant AR" experiences.
Scaling to Global Trajectories
The implications extend beyond single-pair matching. Tian Yi Lim and Hermann Blum use these monocular priors to replace heavy Bundle Adjustment in SLAM systems. Their 2GO framework uses two-view geometry and depth priors to generate metric loop closure constraints, allowing for scalable trajectory estimation on long robotic missions without the computational burden of maintaining a massive 3D landmark database.