Image matching that takes lines seriously

What it does

GlueStick finds correspondences between images using both point and line features in a single deep matcher. Instead of treating keypoints and line segments as separate problems, it reasons about both geometric primitives together. The goal is to improve image matching by leveraging points and lines simultaneously.

The interesting bit

The paper title calls the approach “sticking points and lines together,” and the repository delivers a single joint deep matcher for both primitives. Handling points and lines in one unified model rather than separate stages is the architectural bet here.

Key highlights

• Joint point-and-line matching in a single model, not separate stages
• ICCV 2023 official implementation with a Colab demo notebook
• Training code and framework live in the separate GlueFactory repository
• Built on a SuperPoint backbone
• A DISK-based replacement is in the works to remove the non-commercial SuperPoint dependency

Caveats

• The SuperPoint backbone carries a non-commercial license, so despite the MIT-licensed code, the overall system is not commercially usable
• Training code is absent from this repo; you will need GlueFactory to retrain
• The DISK-based alternative mentioned to resolve licensing is still in progress

Verdict

Worth a look if your image-matching work needs both point and line correspondences in a single model. Skip it if you need a commercially unrestricted, ready-to-ship matcher today.