zju3dv/EfficientLoFTR

What it solves

EfficientLoFTR addresses the trade-off between matching quality and processing speed in local feature matching. It provides semi-dense matching (finding many corresponding points between two images) while achieving speeds comparable to sparse matching methods.

How it works

The system matches local features between image pairs using a semi-dense approach. It offers different configurations to balance performance and speed:

• Model Types: Users can choose between a full model for maximum matching quality or an opt (optimized) model for higher efficiency with minimal loss in quality.
• Precision Modes: It supports multiple precision levels, including full precision (fp32), mixed precision (mp), and half precision (fp16) to further accelerate inference on modern GPUs.
• Integration: The project is integrated into the Hugging Face Transformers library, allowing for streamlined inference using AutoImageProcessor and AutoModel.

Who it’s for

This tool is designed for researchers and developers working in computer vision, specifically those needing efficient image matching for applications like UAV localization, autonomous driving, and 3D reconstruction.

Highlights

• Semi-Dense Matching: Combines the density of feature matching with the speed of sparse methods.
• Flexible Efficiency: Multiple model types and precision settings allow users to tune the balance between speed and accuracy.
• Hugging Face Integration: Easily deployable via the transformers library.
• Broad Application: Specifically trained on MegaDepth for high performance in outdoor scenes.