facebookresearch/silk

What it solves

SiLK tackles the challenge of learning image keypoints — distinct points in an image used for matching and 3D reconstruction — without needing human-labeled data. Traditional methods require expensive manual annotations, while SiLK provides a self-supervised approach that trains models to find keypoints automatically, achieving state-of-the-art results on existing benchmarks.

How it works

SiLK uses a self-supervised framework that learns keypoints from unlabeled images by relying on geometric consistency and visual cues. The training process runs on two GPUs (tested with Tesla V100-SXM2) and takes about 5 hours. The released code includes pre-trained models and supports custom backbones, inference, evaluation, and conversion to TorchScript for deployment. It also integrates with COLMAP for 3D reconstruction workflows.

Who it’s for

Researchers and developers working on computer vision tasks that need reliable keypoint detection and matching, such as image stitching, object tracking, or 3D scene reconstruction — especially those who want a self-supervised method that avoids manual labeling and is easy to adapt.

Highlights

• Self-supervised: no human-labeled keypoint data required.
• Simple and flexible: supports custom backbones and easy integration.
• Competitive results: state-of-the-art performance on standard benchmarks and the IMC2022 challenge.
• Pre-trained models provided, plus tools for evaluation, inference, and TorchScript conversion.
• Includes COLMAP integration for 3D reconstruction pipelines.