cvg/LightGlue

What it solves

LightGlue addresses the need for fast and accurate image matching. It identifies corresponding points between two images, which is a fundamental task in computer vision, while significantly reducing the computational time required compared to previous methods.

How it works

LightGlue is a deep neural network that uses an adaptive mechanism to process image pairs. It dynamically adjusts its depth (how many layers it uses) and width (how many keypoints it considers) based on the complexity of the image pair. For "easy" pairs, it finishes quickly; for "difficult" ones, it optimizes the computation to maintain accuracy without unnecessary overhead. It supports various feature extractors like SuperPoint, DISK, ALIKED, and SIFT.

Who it’s for

It is designed for developers and researchers working on computer vision tasks such as visual localization or Structure-from-Motion who need high-performance feature matching that can run efficiently on both GPUs and CPUs.

Highlights

• Offers a 4-10x speedup over previous state-of-the-art matchers like SuperGlue.
• Features adaptive pruning, which automatically skips unnecessary calculations for simpler image pairs.
• Supports FlashAttention and PyTorch compilation for maximum performance.
• Integrates directly with Hugging Face Transformers for easy implementation in Python.
• Provides flexible configuration to trade off between speed and accuracy based on specific hardware needs.