When neural networks process spatial or geometric data—such as 3D point clouds, geographic maps, 2D sketches, or routing graphs—they frequently struggle with changes in scale, translation, and density. If a model learns to recognize shapes or optimize routes using absolute global coordinates, it inadvertently couples the intrinsic structure of the data with its extrinsic position and magnitude.
Local Coordinate Normalization resolves this fundamental bottleneck. Instead of feeding raw global coordinates into a network, this paradigm transforms spatial data into localized, canonical coordinate systems. By shifting the frame of reference to a local origin (such as a centroid) and scaling by a local metric (such as a neighborhood radius or bounding box), models can isolate structural geometry from global topology. This simple geometric regularizer allows neural networks to achieve scale invariance, stabilize training, and generalize to massive, unseen problem sizes without retraining.
The Intuition: Disentangling Structure from Position
To understand why absolute coordinates fail, consider a 3D point cloud representing an object. If the object moves across the scene or doubles in size, its absolute coordinate values change entirely. A standard neural network must dedicate capacity to learning that the same object can exist at any scale or position in space.
Local coordinate normalization strips away this coordinate bias by enforcing a localized mathematical transformation:
$$\mathbf{x}_{\text{local}} = \frac{\mathbf{x} - \mathbf{\mu}}{\sigma}$$
Where $\mathbf{\mu}$ represents a local anchor point (like a centroid or local mean) and $\sigma$ represents a local scale factor (like a search radius or variance).
In 3D deep learning, this transformation acts as a powerful proxy for explicit geometric modeling. For instance, in General Point Model Pretraining with Autoencoding and Autoregressive, Zhangyang Gao et al. demonstrate that normalizing local point cloud patches relative to their centroids isolates local structural patterns from global spatial positions. This allows the model to learn geometry independently of absolute position in 3D space.
Similarly, Xu Ma et al. show in Rethinking Network Design and Local Geometry in Point Cloud: A Simple Residual MLP Framework that centering and scaling local point clusters removes the influence of absolute coordinates entirely. This lightweight affine transformation forces the network to prioritize relative spatial relationships, stabilizing training and allowing simple multi-layer perceptrons (MLPs) to match the performance of highly complex geometric kernels. The authors of LatticeNet: Fast Point Cloud Segmentation Using Permutohedral Lattices leverage a similar intuition, subtracting the local mean of point positions within a neighborhood to force the feature extractor to focus strictly on the relative distribution and shape of the point cluster.
Generalization Across Orders of Magnitude
The power of local coordinate normalization becomes exceptionally clear in combinatorial optimization problems, such as the Traveling Salesman Problem (TSP). If a neural solver is trained on graphs containing 100 nodes within a unit square, it will completely fail when deployed on a graph with 10,000 nodes, because the global node density and spatial distribution are fundamentally different.
By applying a divide-and-conquer strategy paired with local normalization, networks can solve problems orders of magnitude larger than those they were trained on:
- Sub-graph Partitioning: A massive global problem is broken down into small, localized sub-graphs or neighborhoods.
- Canonical Transformation: Each sub-graph undergoes min-max normalization, translation, or rotation to map its coordinates into a standard unit range.
- Zero-Shot Execution: A neural network trained on small instances processes these normalized sub-graphs. Because the local coordinate distribution matches the training distribution, the model remains highly accurate.
[Massive Global Graph] ──> [Extract Sub-graph] ──> [Translate & Scale to Unit Range] ──> [Neural Solver]
In INViT: A Generalizable Routing Problem Solver with Invariant Nested View Transformer, Han Fang et al. utilize local coordinate normalization within neighborhood views to ensure that geometric representations remain consistent for the neural network regardless of the global problem size or node density. This allows a model trained on small-scale instances to generalize to problems two orders of magnitude larger without retraining.
This principle is mirrored in hierarchical frameworks like GLOP, where Haoran Ye et al. apply min-max normalization and rotation to local sub-graphs to create a scale-invariant input distribution for real-time routing solvers. Furthermore, Zhi Zheng et al. demonstrate in UDC: A Unified Neural Divide-and-Conquer Framework for Large-Scale Combinatorial Optimization Problems that normalizing the spatial coordinates and constraints of sub-problems allows high-performance local experts to be applied to massive graphs without requiring computationally expensive retraining on full-scale data.
Canonical Spaces in Vision, Graphics, and Meshes
Beyond point clouds and graphs, mapping inputs into a standardized unit space allows models to learn structural deformation and part identities independently of global context.
Bounding Box and Origin Normalization
In document analysis and object detection, normalizing coordinates relative to a localized proposal box enables scale-invariant shape regression. In SReN: Shape Regression Network for Comic Storyboard Extraction, Zheqi He et al. normalize vertex coordinates relative to the center and dimensions of a proposal bounding box rather than the global image space, simplifying the learning task by focusing the network entirely on local geometric offsets.
For instance segmentation, Zichen Liu et al. show in DANCE : A Deep Attentive Contour Model for Efficient Instance Segmentation that normalizing vertex coordinates relative to an object's bounding box allows a model to learn shape deformation independently of absolute scale, keeping the numerical range of predicted offsets consistent whether the target object is large or small.
This decoupling is also vital for generative stroke editing; Tengjie Li et al. demonstrate in SketchEdit: Editing Freehand Sketches at the Stroke-Level that normalizing individual stroke coordinates to a common origin during encoding allows a model to learn a geometric shape dictionary independent of global position, enabling precise part swapping and editing.
Articulated and Continuous Domains
For complex, non-rigid structures like the human body, global coordinates fail to capture local pose variations. In COAP: Compositional Articulated Occupancy of People, Marko Mihajlovic et al. transform spatial points into local canonical coordinate systems specific to each articulated segment before processing them with a neural decoder, forcing the network to focus on local shape variations rather than global spatial positioning.
This concept extends seamlessly to continuous physical simulations. In UGM2N: An Unsupervised and Generalizable Mesh Movement Network via M-Uniform Loss, Zhichao Wang et al. operate on normalized local geometric features within small mesh patches. This design choice decouples the adaptation logic from the global coordinate system and total mesh size, ensuring that learned node-movement strategies remain effective across entirely different physical domains and complex geometries.
Robustness to Density and Environmental Shifts
In real-world applications, spatial data is rarely uniform. LiDAR scans vary in point density depending on distance from the sensor, and geographical data shifts based on regional topography. Local coordinate normalization acts as a buffer against these environmental variations.
[Raw Points: Variable Density] ──> [Normalize by Local Search Radius] ──> [Consistent Input Distribution]
In PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies, Guocheng Qian et al. demonstrate that normalizing relative point coordinates by the local search radius ensures that the input distribution remains consistent across different stages of a hierarchical network, making the model robust to variations in point density and spatial scale.
To mitigate sensor distance bias in autonomous driving, Guangyao Zhai et al. show in PASS3D: Precise and Accelerated Semantic Segmentation for 3D Point Cloud that transforming individual object clusters into their own local coordinate systems allows a neural network to learn intrinsic object shapes independently of their absolute position or orientation within the global scene.
When dealing with large-scale aerial LiDAR, Jernej Nejc Dougan et al. show in Semantična segmentacija aerolaserskih oblakov točk in centriranje višin globalnih soseščin that neighborhood-based statistical centering (using local metrics like the mean or minimum within a defined radius) makes the model invariant to global shifts in topography.
Ultimately, whether a model is approximating mathematical hypervolumes by mapping arbitrary input sets into a standardized unit space—as shown by Ke Shang et al. in HV-Net: Hypervolume Approximation based on DeepSets—or upsampling surfaces by shifting local neighborhoods to a canonical origin—as shown by Wenbo Zhao et al. in Self-Supervised Arbitrary-Scale Point Clouds Upsampling via Implicit Neural Representation—the underlying mechanism remains identical. By forcing the network to view the world through a localized, normalized lens, researchers decouple intrinsic geometry from extrinsic scale, unlocking unprecedented generalization across spatial domains.
Key Foundational Work
SReN: Shape Regression Network for Comic Storyboard Extraction (2017) — contributed 2 concept elements to this direction.
A Network of Localized Linear Discriminants (1991) — contributed 1 concept elements to this direction.