When algorithms interpret the 3D world—whether reconstructing a room from images, tracking an object's pose, or processing LiDAR sweeps—they are plagued by outliers. Sensor noise, perspective shifts, and matching ambiguities generate erroneous data points that do not correspond to physical surfaces.
If a system blindly feeds these outliers into geometric optimization loops (like centroid calculation or pose fitting), the results skew dramatically. Traditional robust estimators like RANSAC attempt to solve this by iteratively guessing and checking hypotheses, but this becomes computationally expensive and prone to failure when the outlier ratio is high.
A powerful paradigm has emerged to solve this: prior-guided outlier filtering. Instead of relying purely on downstream optimization to average out errors, these methods leverage statistical, structural, or multi-view priors to explicitly identify and discard outliers before final geometric modeling occurs. By cleaning the data early, systems can use fast, closed-form algorithms to achieve highly accurate 3D perception.
Statistical Isolation of Geometric Anomalies
One way to identify outliers is to treat them as statistical anomalies within a point cloud. In a typical 3D scan of an object, valid geometric measurements form dense, coherent structures, whereas noise from monocular depth estimation or sensor artifacts manifests as sparse, scattered points.
Valid Points (Dense Cluster) Outliers (Isolated & Sparse)
* * * * .
* * * * * .
* * * * .
To exploit this structural difference, Yanmin Wu et al. adapted the Isolation Forest algorithm for spatial data in EAO-SLAM and their subsequent object SLAM framework. An Isolation Forest works by recursively partitioning data points using random axis-aligned splits. Because outliers are sparse and distinct from the main cluster, they require significantly fewer recursive partitions to isolate. By pruning points that isolate quickly, the authors prevent noisy depth measurements from distorting the calculated scale and center of 3D objects.
A related approach relies on residual errors during shape alignment. When matching a partial scan to a known CAD model, sensor noise can disproportionately pull the alignment away from the true solution. In U-RED, Chenyangguang Zhang et al. introduced a residual-guided retrieval mechanism that calculates displacement vectors between scan points and a target surface. By selectively ignoring a fixed percentage of points with the highest residual magnitudes, the system ensures that shape retrieval remains robust against heavy sensor noise.
Leveraging Local Geometric and Physical Priors
Physical objects follow predictable geometric patterns; they are composed of smooth surfaces, sharp boundaries, or continuous curves. Outliers, by contrast, frequently violate these local geometric constraints.
Early work by Alireza Bab-Hadiashar et al. established this by using physical constraints—specifically bending and twisting energy—to model curved surfaces. Their algorithm calculates an estimated noise scale and rejects any points that deviate significantly from the fitted geometric model, ensuring that segmentation is not corrupted by stray range data.
In modern depth completion, researchers use pre-trained models or structural references as geometric priors to anchor sparse data:
- Superpixel RANSAC Alignment: In Zero-shot Depth Completion, Hyoseok Lee et al. handle noise in sparse depth datasets by segmenting the scene into superpixels guided by an affine-invariant depth prior. They apply RANSAC-based linear regression locally within each segment. This divide-and-conquer strategy easily spots and removes outlier points that fail to conform to the local geometry predicted by the pre-trained model.
- Reference Map Rectification: In DenseLiDAR, Jiaqi Gu et al. target "penetration points"—a common artifact where distant LiDAR returns erroneously slip through foreground objects due to perspective shifts. By comparing the sparse input against a coarse, dense reference map, the system identifies and discards points that deviate significantly from their local neighborhood, preventing long-tail errors.
- Exploiting Artifact Behavior: Rather than treating noise as purely random, Yiming Zhao et al. observed that incorrectly mapped LiDAR points exhibit predictable physical behaviors: they tend to shift relative to correct points, possess larger depth values, and form artificial high-density regions. They designed an outlier removal algorithm that explicitly targets these behavioral signatures to clean sparse depth data before surface modeling.
Multi-View and Correspondence Consistency
When multiple viewpoints or point correspondences are available, geometric consistency acts as a strict filter. A true 3D point must project accurately across all views, and a valid correspondence must align with the global transformation of the object.
In Training-Free Instance-Aware 3D Scene Reconstruction, Jiatong Xia et al. eliminate point cloud outliers by projecting 3D points back into multiple 2D image views. If a point creates a depth discrepancy or geometric inconsistency when warped across views, it is flagged as an outlier and eliminated. This multi-view verification ensures a clean, unified 3D representation without requiring per-scene optimization.
View 1 Depth Prior <====== 3D Point ======> View 2 Depth Prior
||
Does it match both views?
/ \
[Yes] [No]
|| ||
Keep Valid Point Discard Outlier
This concept is pushed even further in category-level object pose estimation. Instead of relying on iterative consensus loops to find the correct pose out of a sea of noisy matches, the SpotPose framework by Huan Ren et al. explicitly predicts an outlier score for each point correspondence.
By evaluating the quality of the correspondences early, points exceeding a strict threshold are discarded before the pose calculation begins. Removing these noisy data points early allows the system to utilize fast, closed-form mathematical algorithms for the final geometric alignment, achieving both higher accuracy and superior computational efficiency in heavily occluded environments.
Key Foundational Work
EAO-SLAM: Monocular Semi-Dense Object SLAM Based on Ensemble Data Association (2020) — contributed 2 concept elements to this direction.
Model Selection for Range Segmentation of Curved Objects (2004) — contributed 1 concept elements to this direction.