When 3D sensors—such as LiDAR, RGB-D cameras, or photogrammetric pipelines—capture the physical world, they consistently degrade at object boundaries. Whether due to the physical width of a laser beam straddling an edge or the challenges of stereoscopic triangulation across vast depth changes, sensors produce systemic geometric artifacts. These manifest as "flying pixels" suspended in mid-air, "ghost edges" connecting foreground and background elements, and noisy clusters at extreme viewing angles.
Rather than relying purely on statistical outlier removal, a prominent research direction leverages geometric filtering constraints derived from the physics of the sensor and the spatial structure of the scene. By analyzing depth discontinuities, perspective-ray relationships, and local angular neighborhoods, these techniques isolate true surface geometry from edge-induced noise.
The Anatomy of Edge Noise: Why Discontinuities Fail
To understand why geometric edge filtering is necessary, consider what happens when a sensor grid samples a sharp depth discontinuity. In a mesh reconstruction, a pixel on the edge of a foreground object often gets incorrectly connected to the background. This creates highly elongated, stretched triangles along the sensor's line of sight.
Sensor
o ------------> [Foreground Edge] \
------------> \ <-- Elongated "Ghost" Edge
------------> [Background Plane] /
To resolve this, researchers target the geometric properties of these artifacts:
- Projected Edge and Triangle Stripping: In view interpolation, Carsten Haubold et al. demonstrate that processing depth maps with a geometry shader that discards triangles with large projected edges effectively cleans up depth discontinuities in real-world datasets (Ambient point clouds for view interpolation). Similarly, in medical body registration, Tolga-Can Çallar et al. employ geometric filtering of elongated triangles to prune the stretched polygons created by rapid depth changes at the edges of a body (Efficient Body Registration Using Single-View Range Imaging and Generic Shape Templates).
- Depth Gradient Analysis: Instead of evaluating triangles, Naoto Ienaga et al. analyze depth variations directly along 3D line segments using a grid-based sampling approach (Removing Redundant Line Segments of 3D Model for Camera Pose Estimation). If the depth gradient along a segment exceeds a threshold relative to its length, it is identified as a "jump edge" or "ghost" artifact and removed.
- Pre-Fusion Filtering: Rather than cleaning a completed mesh, Patrick Stotko et al. apply depth discontinuity filtering early in the pipeline, discarding samples on sharp edges or near missing data before volumetric fusion occurs (Efficient 3D Reconstruction and Streaming for Group-Scale Multi-Client Live Telepresence). This prevents "flying pixels" from ever entering the reconstruction.
- Boundary-Guided Localization: When the region of interest is semantically known, noise reduction can be explicitly steered. Han Joo Chae et al. show that focusing noise reduction specifically on the perimeter of a segmented area yields high-fidelity boundaries required for subsequent geometric processing and fabrication (Generating 3D Bio-Printable Patches Using Wound Segmentation and Reconstruction to Treat Diabetic Foot Ulcers).
Ray-Surface Geometry and Grazing Angles
A second major source of noise stems from the geometric relationship between the sensor's perspective ray and the orientation of the target surface. When a camera or scanner views a surface at a "grazing angle"—where the line of sight is nearly parallel to the surface itself—minor errors in pixel matching or sensor calibration translate into massive spatial errors in 3D.
In oblique aerial photogrammetry, this geometric vulnerability is highly pronounced at building facades and roof edges. To counteract this, Francesco Nex et al. introduce a filter based on the angular relationship between the camera's perspective ray and the local surface normal (Building Footprints Extraction from Oblique Imagery). By discarding points where the ray aligns too closely with the surface normal at boundaries, they eliminate mixed pixels.
Expanding on this framework, Ewelina Rupnik et al. demonstrate that removing points where the camera's line of sight is nearly parallel to the local surface normal systematically purifies dense matching point clouds, targeting the exact artifacts that occur at grazing angles (OBLIQUE MULTI-CAMERA SYSTEMS – ORIENTATION AND DENSE MATCHING ISSUES).
This perspective-ray constraint is also vital when combining heterogeneous sensors. For instance, in autonomous driving pipelines where LiDAR point clouds are projected onto 2D camera images, spatial offsets (parallax) cause background points to erroneously project into foreground masks. As shown by Min Dou et al., enforcing local geometric constraints along these projection rays allows the system to identify and prune parallax-induced errors, ensuring clean 3D segmentation (ZOPP: A Framework of Zero-shot Offboard Panoptic Perception for Autonomous Driving).
Angular Neighborhoods and Sensor-Level Denoising
At the hardware level, sensors like LiDAR suffer from the veiling effect: when a laser pulse partially strikes a foreground edge, the beam splits, returning a mixed signal that creates a phantom point in the empty space between the two surfaces.
LiDAR Pulse =====> [Foreground Edge] --> Partial Return
=====> |
=====> v --> Phantom Point (Veiling Effect)
[Background Wall] --> Partial Return
To catch these phantom points, the local angular distribution of neighboring points can be analyzed. Donghai Yang et al. develop a geometric filter that computes the angular relationships between a point and its nearest neighbors (Development of the high angular resolution 360° LiDAR based on scanning MEMS mirror). If the sum of angles to its neighbors exceeds a specific threshold, it indicates a lack of spatial consistency, identifying the point as a split-beam phantom distortion.
Alternatively, the raw structure of the sensor's data collection can be leveraged. Yiting Liu et al. introduce a "ring band filter" for LiDAR preprocessing (Laser Radar Data Registration Algorithm Based on DBSCAN Clustering). By analyzing a point's neighborhood using concentric inner and outer bands, the filter isolates slope outliers, allowing the system to distinguish between true structural corner features and edge noise.
Macro-Geometric and Model-Driven Constraints
When local neighborhood analysis is insufficient, macro-level geometric assumptions—ranging from local curvature offsets to global structural models—can act as powerful filters.
Geometric Offsets and Primitives
On a mathematical level, high-frequency noise can be smoothed by treating shape offsets as a low-pass filter. David Cohen-Steiner et al. establish a rigorous framework proving that defining a shape as the set of points within a specific distance from the input data stabilizes curvature estimation for noisy, non-smooth point clouds (Stability of Curvature Measures).
On an algorithmic level, if the target geometry conforms to known shapes, fitting primitives provides an absolute reference for denoising. Qingguo Zhang et al. isolate specific surface features within tunnels by fitting cylinders to the global data and filtering out points based on their radial distance, separating structural components from transient interference (Channel Extraction and Geometric Parameters Measurement Based on Point Clouds).
Global Reference Models
In industrial and architectural settings, an existing design model can serve as a geometric anchor. Yan Zhou et al. demonstrate that by selecting only the outermost points relative to a tunnel's predefined design geometry, a system can filter out transient clutter like machinery and personnel without requiring complex object classification (EVALUATION OF TUNNEL EXCAVATION COMBINING TERRESTRIAL LASER SCANNING POINT CLOUDS AND DESIGN MODELS). This works because the true structural boundary represents the furthest physical extent from the tunnel center.
[Design Model Boundary]
|
v
* * * * * <-- Outermost Points Retained
[Clutter/Worker]
o o o <-- Internal Points Filtered Out
For large-scale environments where no design model exists, bounding the geometry globally prevents infinite depth errors. As shown by Qingling Chang et al., explicitly masking out infinite depth regions (such as the sky or far horizons) ensures that surface reconstruction algorithms focus entirely on reachable, valid geometry, resulting in cleaner architectural meshes (3D Reconstruction With Spherical Cameras).
Finally, fitting global primitives can overcome the hardware limits of the sensor itself. Ye Wang et al. show that by fitting a single geometric plane to thousands of points, a system effectively averages out stochastic depth noise, allowing calibration accuracies to exceed the native precision of the depth hardware (PlaneHEC: Efficient Hand-Eye Calibration for Multi-view Robotic Arm via Any Point Cloud Plane Detection).
Key Foundational Work
Ambient point clouds for view interpolation (2010) — contributed 1 concept elements to this direction.
Stability of Curvature Measures (2008) — contributed 1 concept elements to this direction.