SuperPoint: Self-Supervised Interest Point Detection and Description

By Daniel DeTone, Tomasz Malisiewicz, Andrew Rabinovich. 2017

3,557 citations recorded by the source corpus.

This paper presents a self-supervised framework for training interest point detectors and descriptors suitable for a large number of multiple-view geometry problems in computer vision. As opposed to patch-based neural networks, our fully-convolutional model operates on full-sized images and jointly computes pixel-level interest point locations and associated descriptors in one forward pass. We introduce Homographic Adaptation, a multi-scale, multi-homography approach for boosting interest point detection repeatability and performing cross-domain adaptation (e.g., synthetic-to-real). Our model, when trained on the MS-COCO generic image dataset using Homographic Adaptation, is able to repeatedly detect a much richer set of interest points than the initial pre-adapted deep model and any other traditional corner detector. The final system gives rise to state-of-the-art homography estimation results on HPatches when compared to LIFT, SIFT and ORB.

Subfields: Aerospace Engineering, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Aerospace Engineering

In Computer Vision and Pattern Recognition

Citations in the corpus

Ego4D: Around the World in 3,000 Hours of Egocentric Video (2021, background): “[50] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich.”
Lifelong Graph Learning (2020, positive): “For example, SuperPoint [9] introduced a self-supervised framework for extracting interest point detectors and descriptors.”
CoTracker3: Simpler and Better Point Tracking by Pseudo-Labeling Real Videos (2025, positive): “5, we have explored alternative point sampling methods, including LightGlue [20], Super-Point [5], and DISK [36].”
ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformer (2022, methodology): “As is shown in previous works [10,8,9,12,28,27,26,29], CNN exhibits strong capability to capture local context and generates high-level features, which can be directly used to perform nearest neighbour matching.”
Learning Target Candidate Association to Keep Track of What Not to Track (2021, methodology): “Therefore, we introduce the concept of dustbins, which is commonly used for graph matching [38, 18] to actively handle the non-matching vertices.”
D2-Net: A Trainable CNN for Joint Description and Detection of Local Features (2019, methodology): “These methods follow a detect-then-describe approach that first applies a feature detector [7,13,19,29,31,33,49,64] to identify a set of keypoints or interest points.”
Visual Camera Re-Localization From RGB and RGB-D Images Using DSAC (2020, background): “Learning-based sparse feature pipelines [41], [42], [43], [44] might ultimately be able to overcome these issues, but currently it is an open research question whether learned sparse features consistently exceed the capabilities of their hand-crafted predecessors [45], [46].”
LF-Net: Learning Local Features from Images (2018, methodology): “Note however that the self-supervised training strategy used by SuperPoint can also be applied to LF-Net to boost performance.”
Key.Net: Keypoint Detection by Handcrafted and Learned CNN Filters (2019, methodology): “A number of related works have focused their objective function to address this issue, although they were based either on local patches [9, 10] or global map regression loss [13, 14, 11].”
DUSt3R: Geometric 3D Vision Made Easy (2023, neutral): “…modern approaches for 3D reconstruction consists in assembling the fruits of decades of advances in various sub-fields such as keypoint detection [26, 28, 63, 97] and matching [10, 60, 100, 120], robust estimation [3, 10, 181], Structure-from-Motion (SfM) and Bundle Adjustment (BA) [20, 59,…”
COTR: Correspondence Transformer for Matching Across Images (2021, background): “They rely on local features [34, 74, 44, 13] Figure 1.”
DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras (2021, methodology): “Stereo OV 2 SLAM [17] 0.510 0.182 VOLDOR [28] + COLMAP [41] 0.440 0.177 SuperGlue [39] + SuperPoint [13] + COLMAP [41] In the monocular setting, we achieve an average ATE of 2.2cm, reducing error by 82% among methods with zero failures, and by 43% over ORB-SLAM3 when only comparing sequences where O”
D3Feat: Joint Learning of Dense Detection and Description of 3D Local Features (2020, background): “In 2D image matching, several works have tackled the keypoint detection and description problems jointly [35, 6, 20, 24, 3, 26, 19].”
Emergent Correspondence from Image Diffusion (2023, methodology): “Following SuperPoint [17] and CAPS [87], we extract a maximum of 1,000 keypoints from each image, and use RANSAC to estimate the homography from mutual nearest neighbor matches.”
iNeRF: Inverting Neural Radiance Fields for Pose Estimation (2020, methodology): “re have seen real posed images, rather than only synthetic posed images. rotation augmentation is crucial because feature descriptor is not rotation-equivariant. Speciﬁcally, we ﬁrst apply SuperPoint [6] for feature extraction and SuperGlue [34] for correspondence matching. Then, a sparse 3D model o”