LoMa: Local Models & Methodologies

• LoMa is a family of approaches that harness local data structures for tasks in computer vision, memory compression, and semantic mapping.
• It integrates advanced transformer architectures, state-space models, and geometric augmentation techniques to enhance accuracy and efficiency.
• LoMa variants, from pixel-level forgery detection to LiDAR odometry, consistently deliver superior performance on challenging benchmarks.

LoMa refers to a family of independently developed models, methods, and frameworks across computer vision, machine learning, data augmentation, and LLM memory optimization. The term is associated with diverse methodologies ranging from local feature matching for 3D vision, lossless memory compression in transformers, pixel-level tampering localization in images, geometric data augmentation, local manifold-based classification, vision-language occupancy prediction, and multiple variants within LiDAR odometry and mapping. This article comprehensively surveys the principal LoMa (and LOMA, LOMA, LOAM) frameworks as documented in the arXiv literature.

1. Local Feature Matching: LoMa and Its Scaling Principles

LoMa, as introduced in "LoMa: Local Feature Matching Revisited" (Nordström et al., 6 Apr 2026), represents a new generation of sparse feature matchers for 3D computer vision. LoMa reinstates the detect–describe–match paradigm by integrating large-scale, diverse training mixtures (17 3D and synthetic datasets), state-of-the-art training techniques (dual-softmax, layerwise supervision, AdamW), and expanded transformer-based architectures (B, L, G variants). Unlike traditional sparse matchers, LoMa is trained on orders of magnitude more data and leverages extensive compute, leading to substantial gains on both conventional and novel, highly challenging benchmarks (notably "HardMatch," a new suite of difficult, manually-annotated correspondence pairs).

LoMa employs the DeDoDe descriptor (U-shaped CNN outputs), processed through a multi-layered transformer matcher extending the LightGlue architecture. Final correspondences are assigned via dual-softmax, ensuring mutual consistency. Ablation and scaling studies demonstrate that increasing data, model capacity, and compute all yield measurable improvements in mean average accuracy (mAA). On benchmarks such as HardMatch, WxBS, InLoc, RUBIK, and IMC 2022, LoMa-G achieves mAA and AUC gains up to +29.5 over ALIKED+LightGlue (Nordström et al., 6 Apr 2026).

2. Lossless Compressed Memory Attention in Transformers

LoMA (Lossless Compressed Memory Attention) (Wang et al., 2024) addresses memory and computational bottlenecks in autoregressive transformers for long-context language modeling. Standard models accumulate quadratic GPU memory due to the expanding key-value (KV) cache. LoMA introduces a training scheme that enables the compressed replacement of $c\cdot t$ KV pairs with $t$ special "memory" tokens, occurring periodically during generation.

LoMA's approach is fundamentally lossless within the compression window, as evidenced by repetition loss over a supervised "reproduction zone"; each memory token is trained to recall the original context tokens with near-perfect fidelity for reasonable compression ratios (e.g., $c=4$, $t=8$). The training pipeline implements specific attention masks that enforce bidirectional communication among memory tokens and restrict repetition tokens to attend only to memory. Empirically, LoMA on LLaMA-2-7B reduces compute by up to 74.2% and GPU memory by 800% (at $c=8$), outperforming sparse eviction and prompt-gisting approaches while remaining architecturally simple to integrate (Wang et al., 2024).

3. Image Forgery Localization with State-Space Models

LoMa as formulated in "Image Forgery Localization with State Space Models" (Lou et al., 2024) operationalizes a fully state-space model (SSM)-centric pipeline for pixel-level image tampering localization. Classic CNN and Transformer-based IFL systems lack either sufficient receptive field or incur steep $O(N^2)$ computation. LoMa's pipeline begins with an atrous selective scan to linearize image patches into multi-directional sequences, which are then modeled across four principal directions using selective SSM (Mamba) blocks. This architecture allows global pixel dependency modeling in $O(N)$ time.

A MobileNetV2-derived auxiliary convolutional branch is incorporated for high-resolution, local boundary precision. Final predictions are fused through an MLP-based decoder. LoMa exhibits superior F1 and IoU across six datasets (average IoU: 56.5%) and demonstrates strong robustness to JPEG compression, blur, and resizing, substantiating the practical benefits of SSM-based representations for forensic image analysis (Lou et al., 2024).

4. Local Magnification Data and Feature Augmentation

LOMA in "Local Magnification for Data and Feature Augmentation" (He et al., 2022) extends the data augmentation portfolio with local, non-uniform geometric transformations. Rather than global transforms, LOMA randomly magnifies a small region in the image (ellipse or rhombus shape), redistributing pixels via spatial mapping to induce local scale transformations without obscuring object identity. In practice, this process is parameterized by probabilities, random center, radius, compression ratios, and the selected geometric shape.

LOMA is further extended to intermediate convolutional features (LOMA_IF) and feature map offsets (FO), boosting robustness to spatial perturbations. Empirical results on CIFAR and ImageNet show up to 0.8% accuracy gains over strong baselines when LOMA is applied jointly at image and feature levels. LOMA is additive to intensity-based and mixup-style augmentations, theoretically justified by its coverage of local manifold perturbations and empirically validated by ablations (He et al., 2022).

5. Local Manifold Approximation (LOMA) for Classification

LOcal Manifold Approximation (LOMA) defines a principled, geometry-driven framework for nonparametric classification under complex class support structure (Li et al., 2019). For an input $x$, LOMA fits a parametric manifold (sphere, tangent plane, or quadratic surface) to a local neighborhood of each class, assigns $x$ to the class whose manifold is closest in Euclidean distance, and leverages local geometric properties for decision making.

The SPA (Spherical Approximation) variant fits spheres of intrinsic dimension $p$ using projected principal components and closed-form spherical fitting. The method offers consistency guarantees under both noiseless and noisy data distributions provided class manifolds are non-overlapping or have small intersection. SPA-LOMA achieves high accuracy, especially in the limited sample regime and for data with nonlinear, intersecting supports (e.g., interlocking curves, handwritten digits), consistently outperforming SVMs, random forests, and CNNs on various benchmarks (Li et al., 2019).

6. Vision-Language Semantic Occupancy Networks

LOMA as "Language-assisted Semantic Occupancy Network via Triplane Mamba" (Cui et al., 2024) integrates vision-language priors for large-scale 3D semantic occupancy prediction. The approach constructs 3D features from both visual (image, depth) and language (object class tokens) modalities, lifts these into voxel grids, and fuses them through a tri-plane Mamba (state-space model) block. The tri-plane design performs efficient linear-complexity global 3D fusion by projecting volumetric features to canonical planes, applying SSM convolutions, and back-projecting to 3D.

A vision-language-aware scene generator injects class-conditional geometric and semantic context, while training losses combine semantic classification (cross-entropy) and occupancy IoU. On SemanticKITTI and SSCBench-KITTI360, LOMA attains new state-of-the-art IoU and mIoU, outperforming strong baselines such as VoxFormer and OccFormer (Cui et al., 2024).

7. LOAM Variants in LiDAR Odometry and Mapping

"LOAM" (LiDAR Odometry and Mapping) and its derivatives, including P³-LOAM (Li et al., 2020), CAR-LOAM (Lu et al., 24 Feb 2025), InTEn-LOAM, and Splat-LOAM (Giacomini et al., 21 Mar 2025), are widely adopted pipelines in mobile robotics and autonomous navigation. They share a multi-stage structure: feature extraction (typically edges and planes), scan-to-map registration (often via ICP), and global optimization (factor-graph or keyframe-based).

Innovations across variants include:

• P³-LOAM: Tight GNSS-LiDAR fusion with explicit SVD-Jacobian-based ICP covariance estimation and LiDAR-augmented GNSS RAIM for outlier tolerance; achieves sub-4 m global accuracy in urban canyons (Li et al., 2020).
• CAR-LOAM: Integrates RGB camera cues to colorize LiDAR points, applying robust pose optimization with Welsch and color-difference-based weighting, improving localization in low-geometric-context environments (Lu et al., 24 Feb 2025).
• Splat-LOAM: Replaces classical point-cloud mapping with compact, differentiable Gaussian splats, enabling highly efficient, real-time, memory-bounded mapping and state-of-the-art F-scores on benchmark datasets (Giacomini et al., 21 Mar 2025).

These architectures pursue high-precision, memory-efficient, and robust mapping, each adding distinctive algorithmic advantages targeting different field scenarios.

---

In summary, LoMa/LOMA/LOAM encompasses a diverse set of algorithmic innovations exploiting local structure (either in the data, representation, or model context) for superior performance in feature matching, data augmentation, classification, forgery localization, semantic 3D occupancy, and LiDAR-based navigation. Each instantiation is best understood within its specific methodological and application context, but all share the guiding principle of leveraging local or structured information for robust, scalable performance across a range of core machine learning problems.