davnords/LoMa

[ECCV 2026 Oral] LoMa: Local Feature Matching Revisited

LoMa is a family of local feature matchers — algorithms that find corresponding points between two images of the same scene. It's designed as a faster, more robust alternative to existing matchers like LightGlue and RoMa.

What it does: Given two images (e.g., photos of a building from different angles), LoMa identifies which pixels in each image correspond to the same physical point in the world. This is a core building block for 3D reconstruction (Structure-from-Motion), visual localization, and robotics navigation.

Key properties:

  • Speed + accuracy: It's fast like LightGlue but significantly more robust and accurate, even outperforming the heavier RoMa on benchmarks.
  • Drop-in replacement: Because it uses local keypoint descriptions, it can replace existing matchers in SfM and visual localization pipelines without changing the rest of the system.
  • Multiple model sizes: LoMa-B (same size as LightGlue), LoMa-B128, LoMaL, and LoMa-G (heavier, more accurate) — letting users trade speed for precision.
  • Rotation invariance: A variant handles image rotation, a common challenge in real-world matching.

How to use: Simple Python API — load a pretrained model (LoMaB()) and run it on image pairs. The repo includes a demo script and pretrained weights.

Provenance: Academic research from Chalmers, Linköping, Amsterdam, and Lund universities, with papers at ECCV 2026 (main) and CVPR 2026 Workshops (rotation invariance).

Related

  • Dispatch
    Using LoRA for Efficient Stable Diffusion Fine-Tuning

    Hugging Face has integrated Low-Rank Adaptation (LoRA) into the diffusers library, enabling faster Stable Diffusion fine-tuning with significantly lower VRAM requirements and model weights as small as 3 MB.

  • Dispatch
    SDXL and Stable Diffusion Fast Inference with Latent Consistency LoRAs

    Hugging Face introduces LCM LoRAs, a method to enable high-quality image generation in 4 to 8 steps for SDXL and Stable Diffusion models, significantly reducing inference time.

  • Project
    bcmi/libcom

    An image composition toolbox that provides a unified set of tools for realistic object insertion, including harmonization, shadow generation, and placement assessment.

  • Dispatch
    llama.cpp and llama.app: Local LLM Inference and Ecosystem

    llama.cpp and its official home llama.app provide high-performance, hardware-agnostic local LLM inference, now featuring a simplified 'llama serve' command and integration with local coding agents like Pi.

  • Project
    OpenStitching/stitching

    A Python package for fast and robust image stitching that creates panoramas from overlapping images using OpenCV.