Superpoint

SuperPoint is a computer vision method for detecting and describing interest points in images. Introduced in 2018 by Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich, SuperPoint uses a single fully convolutional neural network to jointly predict a keypoint heatmap and a dense descriptor map for an input image. The network shares an encoder and has two heads: a detector that outputs a probability map of likely keypoints and a descriptor head that outputs a per-pixel descriptor (typical length 256). Keypoints are extracted by applying non-maximum suppression to the detector heatmap, and their corresponding descriptors are obtained by sampling from the descriptor map. Matching points between images is performed by nearest-neighbor search in descriptor space, often with a mutual-consistency check.

Training combines synthetic data with a self-supervised real-image signal. Synthetic data provides reliable ground-truth keypoints and correspondences, while a technique known as homographic adaptation applies random projective transformations to real images to generate additional training views, encouraging detector stability across viewpoints.

Applications and impact: SuperPoint provides an end-to-end learned alternative to traditional handcrafted features and has been used as a robust feature extractor for image matching, structure-from-motion, and SLAM pipelines. Open-source implementations in common deep learning frameworks have facilitated adoption and further research, and the approach has inspired subsequent learned detectors and descriptors.

Limitations: While effective in many scenarios, SuperPoint's performance can be sensitive to training data, may struggle in extremely low-texture regions, and can require substantial computational resources for training and inference.