Vanishing points, planes and groups

Beyond points and lines, LIMAP reconstructs groups: structured primitives that many features share. A group is a vanishing point, a plane, or one of the parametric primitives (sphere, cylinder, ellipsoid, cuboid, cone). The points and lines associated with a group are constrained to it during bundle adjustment, which is what makes the reconstruction holistic.

The detectors live under limap.image.groups and follow the same registry pattern as the line detectors in Line detection, description and matching.

Vanishing point detection

Vanishing points are estimated from the 2D lines detected on an image:

import limap.image.groups.vplib as vplib

# lines: list[limap.geometry.Line2d] for one image (see line2d)
detector = vplib.get_vp_detector("jlinkage", vplib.DetectorOptions())
vpresult = detector.detect_vp(lines)

JLinkage and Progressive-X (separate installation needed) are supported.

Plane detection

Plane segmentation is monocular: it runs on a single image, with neither depth nor pose required.

from pathlib import Path
import limap.image.groups.planelib as planelib

detector = planelib.get_plane_detector(
    "pxwplanar", planelib.DetectorOptions()
)
plane_mask = detector.detect_plane_mask(Path("example.png"))

The detector is pxwplanar, pulled in by requirements.txt; its weights are downloaded from Hugging Face on first use. Segmentation parameters are exposed through planelib.PxwPlanarOptions.

Over many images

The batch helpers mirror the line ones and write their results into the structure database:

• limap.image.groups.vp_detection()

• limap.image.groups.plane_detection()

• limap.image.groups.group_description()

In the mapping pipelines, groups are switched on with the use_groups option (enabled by default; see cfgs/structure_triangulation/default.yaml, and default_cpu.yaml for a configuration that disables them).