SuperPoint
lastmod 2026-11-10created 2019-11-073 edits
SuperPoint is a self-supervised neural network that detects interest points in an image and computes their descriptors at the same time. A detector pre-trained on synthetic shapes is trained on real images by applying homography transformations and generating its own labels, so no data with human-annotated interest points is needed. A single network performs both detection and descriptor extraction, replacing classical methods such as SIFT, and it is used as a preprocessing stage in SLAM and image matching pipelines.
Efficient neural feature detector and descriptor.
Documentation
- SuperPoint - Self-Supervised Interest Point Detection and Description
- https://arxiv.org/pdf/1712.07629.pdf