Back

SuperPoint: Self-Supervised Interest Point Detection and Description

Daniel DeToneTomasz MalisiewiczAndrew Rabinovich

article2017CVPR3,687 citations

Proposes a self-supervised fully convolutional network that simultaneously detects keypoints and extracts descriptors in a single forward pass, using homographic adaptation to train on real images without human annotations while outperforming traditional methods like SIFT and ORB.

Listen

SuperPoint introduces a self-supervised neural network that jointly detects interest points and computes their descriptors directly from full-resolution images in a single forward pass. The work addresses a long-standing bottleneck in geometric computer vision: reliable extraction of repeatable 2-D points from raw images under varying illumination and viewpoint, which underpins tasks such as SLAM, structure-from-motion, and image matching. Traditional hand-engineered detectors such as SIFT, FAST, and Harris remain widely used because supervised learning has been hampered by the lack of unambiguous ground-truth labels for interest points in real photographs.

The authors first pre-train a base detector called MagicPoint on a large synthetic dataset of simple geometric shapes whose corner locations are unambiguous. They then apply a novel procedure, Homographic Adaptation, that warps each unlabeled real image through many random homographies, aggregates the detector responses, and produces high-quality pseudo-ground-truth labels. These labels are used to train the full SuperPoint model on the MS-COCO dataset; the architecture shares a VGG-style encoder between a point-detection head and a descriptor head, enabling real-time operation at approximately 70 frames per second on 480-by-640 images.

On the HPatches benchmark, SuperPoint achieves the highest repeatability under illumination change and matches or exceeds classical detectors under viewpoint change. When the detected points and descriptors are used for homography estimation, the system outperforms LIFT and ORB and performs comparably to SIFT while producing denser, more evenly distributed correspondences. The learned descriptors also score higher on nearest-neighbor matching metrics than the hand-crafted alternatives.

These results indicate that a single, efficient network can replace the traditional detect-then-describe pipeline for many geometric tasks, offering both speed and improved robustness without requiring expensive 3-D supervision. The approach therefore lowers the barrier to deploying learned features in robotics and augmented-reality pipelines that must run in real time.

The principal limitations are that performance still degrades under extreme in-plane rotations not well represented in training and that the method relies on homographies, which only approximate general 3-D motion. Further validation on broader outdoor and dynamic scenes, together with integration into complete SLAM or SfM systems, would strengthen confidence in wider deployment.

Magic Leapboostingcomputer visioncorner detectorSuperPoint interest point detectorSelf-supervised interest point detectionLocal descriptors
arXiv: 1712.07629magicleap/SuperPointPretrainedNetwork
  • Paper: Distinctive Image Features from Scale-Invariant Keypoints, David G. Lowe (2004). Reading the foundational SIFT paper provides essential context on scale-invariant feature extraction and keypoint description that SuperPoint builds upon and automates.
  • Paper: Machine Learning for High-Speed Corner Detection, Edward Rosten et al. (2006). This paper establishes the necessity of high-speed corner detection in computer vision, framing the computational bottlenecks that SuperPoint addresses with a single-pass convolutional model.
Cover for SuperPoint: Self-Supervised Interest Point Detection and Description

Abstract

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.

Topics

boostingcomputer visioncorner detectordetection algorithminterest pointsneural networkssiftfeature correspondenceviewpointepipolar

Keywords

SuperPoint interest point detectorSelf-supervised interest point detectionLocal descriptorsMagicPoint detectorInterest point decoderPseudo-ground-truth interest point locationsImage matchingMS-COCO images

Table of Contents

  • 1. Introduction
  • 2. Related Work
  • 3. SuperPoint Architecture
  • 3.1. Shared Encoder
  • 3.2. Interest Point Decoder
  • 3.3. Descriptor Decoder
  • 3.4. Loss Functions
  • 4. Synthetic Pre-Training
  • 4.1. Synthetic Shapes
  • 4.2. MagicPoint
  • 5. Homographic Adaptation
  • 5.1. Formulation
  • 5.2. Choosing Homographies
  • 5.3. Iterative Homographic Adaptation
  • 6. Experimental Details
  • 7. Experiments
  • 7.1. System Runtime
  • 7.2. HPatches Repeatability
  • 7.3. HPatches Homography Estimation
  • 8. Conclusion
  • A. Evaluation Metrics
  • B. Additional Synthetic Shapes Experiments
  • C. Homographic Adaptation Experiment
  • D. Extra Qualitative Examples

Knowls

  1. Knowl 1 — SuperPoint Fully-Convolutional Architecture

    model/method

    SuperPoint is a fully-convolutional neural network architecture designed to jointly predict interest point locations and fixed-length descriptors on full-sized images in a single forward pass.

    The model consists of a shared VGG-style convolutional encoder and two specialized decoder heads:

    1. Shared Encoder: Processes an input image I∈RH×WI \in \mathbb{R}^{H \times W} through eight 3×33 \times 3 convolutional layers (with channel widths 64-64-64-64-128-128-128-128, interspersed with four 2×22 \times 2 max pooling operations with stride 2 every two layers). It produces a reduced spatial feature map B∈RHc×Wc×128B \in \mathbb{R}^{H_c \times W_c \times 128}, where Hc=H/8H_c = H/8 and Wc=W/8W_c = W/8. Each spatial pixel in BB corresponds to an 8×88 \times 8 non-overlapping pixel region ("cell") in the original image.

    2. Interest Point Decoder: Computes dense point detection probabilities without computationally intensive deconvolution layers. It applies a 3×33 \times 3 convolution of 256 units followed by a 1×11 \times 1 convolution of 65 units to generate a tensor X∈RHc×Wc×65\mathcal{X} \in \mathbb{R}^{H_c \times W_c \times 65}. The 65 channels represent the 64 pixel locations within the 8×88 \times 8 cell plus an additional 65th "no interest point" dustbin channel. A channel-wise softmax is applied across the 65 channels, the dustbin channel is removed, and the remaining tensor is reshaped via sub-pixel convolution (depth-to-space) from RHc×Wc×64\mathbb{R}^{H_c \times W_c \times 64} to RH×W\mathbb{R}^{H \times W}, yielding full-resolution interest point probabilities.

    3. Descriptor Decoder: Computes semi-dense descriptor representations to reduce computation and memory. It applies a 3×33 \times 3 convolution of 256 units followed by a 1×11 \times 1 convolution of D=256D = 256 units, producing a semi-dense descriptor tensor D∈RHc×Wc×D\mathcal{D} \in \mathbb{R}^{H_c \times W_c \times D}. At inference, dense or sparse descriptors of dimension DD are obtained across the full resolution via bicubic interpolation followed by L2L_2 normalization to ensure unit length.

  2. Knowl 2 — Homographic Adaptation for Self-Supervised Interest Point Detection

    model/method

    Homographic Adaptation is a self-supervised domain-adaptation procedure designed to boost the repeatability and geometric consistency of an interest point detector fθ(⋅)f_\theta(\cdot) when applied to unlabeled real-world target images.

    An ideal interest point operator should be covariant under projective transformations (homographies) H\mathcal{H}, satisfying:

    Hx=fθ(H(I))  ⟹  x=H−1fθ(H(I))\mathcal{H} x = f_\theta(\mathcal{H}(I)) \implies x = \mathcal{H}^{-1} f_\theta(\mathcal{H}(I))

    where II is an input image and xx is the detected interest point set.

    Because real-world detectors are not perfectly covariant, Homographic Adaptation generates an aggregated pseudo-ground truth response map F^(I;fθ)\hat{F}(I; f_\theta) by averaging the unwarped detections from NhN_h sampled homographies:

    F^(I;fθ)=1Nh∑i=1NhHi−1fθ(Hi(I))\hat{F}(I; f_\theta) = \frac{1}{N_h} \sum_{i=1}^{N_h} \mathcal{H}_i^{-1} f_\theta(\mathcal{H}_i(I))

    Homographies Hi\mathcal{H}_i are generated by composing simpler transformations sampled from truncated normal distributions within pre-defined ranges: translation, scaling, in-plane rotation, and symmetric perspective distortion, preceded by a root center crop to prevent border interpolation artifacts. The first homography is set to identity (H1=I\mathcal{H}_1 = I). Aggregating responses across Nh≈100N_h \approx 100 transforms significantly improves interest point repeatability across varied viewpoints and scales.

  3. Knowl 3 — SuperPoint Joint Training Loss Formulation

    equation

    The joint objective function for training the SuperPoint network balances the interest point detector loss LpL_p and the descriptor correspondence loss LdL_d across a pair of synthetically warped images related by a known homography H\mathcal{H}:

    L(X,X′,D,D′;Y,Y′,S)=Lp(X,Y)+Lp(X′,Y′)+λLd(D,D′,S)L(\mathcal{X}, \mathcal{X}', \mathcal{D}, \mathcal{D}'; Y, Y', S) = L_p(\mathcal{X}, Y) + L_p(\mathcal{X}', Y') + \lambda L_d(\mathcal{D}, \mathcal{D}', S)

    where X,X′∈RHc×Wc×65\mathcal{X}, \mathcal{X}' \in \mathbb{R}^{H_c \times W_c \times 65} are the cell-level interest point logits for image pair (I,I′)(I, I'), Y,Y′Y, Y' are ground-truth cell labels, D,D′∈RHc×Wc×D\mathcal{D}, \mathcal{D}' \in \mathbb{R}^{H_c \times W_c \times D} are semi-dense L2L_2-normalized descriptors, SS is the set of ground-truth pairwise correspondences, and λ=0.0001\lambda = 0.0001.

    The interest point loss LpL_p is a fully-convolutional cross-entropy loss over all Hc×WcH_c \times W_c cells:

    Lp(X,Y)=1HcWc∑h=1Hc∑w=1Wclp(xhw;yhw)L_p(\mathcal{X}, Y) = \frac{1}{H_c W_c} \sum_{h=1}^{H_c} \sum_{w=1}^{W_c} l_p(x_{hw}; y_{hw})

    lp(xhw;y)=−log⁡(exp⁡(xhwy)∑k=165exp⁡(xhwk))l_p(x_{hw}; y) = -\log \left( \frac{\exp(x_{hwy})}{\sum_{k=1}^{65} \exp(x_{hwk})} \right)

    where yhw∈{1,…,65}y_{hw} \in \{1, \dots, 65\} denotes the ground-truth bin index within cell (h,w)(h, w). If multiple points land in the same cell, one is chosen at random; if no point is present, yhw=65y_{hw} = 65 (the dustbin).

    The descriptor loss LdL_d operates on all pairs of descriptor cells (dhw∈D,dh′w′′∈D′)(d_{hw} \in \mathcal{D}, d'_{h'w'} \in \mathcal{D}'):

    Ld(D,D′,S)=1(HcWc)2∑h=1Hc∑w=1Wc∑h′=1Hc∑w′=1Wcld(dhw,dh′w′′;shwh′w′)L_d(\mathcal{D}, \mathcal{D}', S) = \frac{1}{(H_c W_c)^2} \sum_{h=1}^{H_c} \sum_{w=1}^{W_c} \sum_{h'=1}^{H_c} \sum_{w'=1}^{W_c} l_d(d_{hw}, d'_{h'w'}; s_{hwh'w'})

    where correspondence indicator shwh′w′∈{0,1}s_{hwh'w'} \in \{0, 1\} is 1 if ∥Hphw−ph′w′∥≤8\|\mathcal{H} p_{hw} - p_{h'w'}\| \le 8 pixels (with phwp_{hw} denoting cell center coordinates) and 0 otherwise. The hinge loss is defined with positive margin mp=1m_p = 1, negative margin mn=0.2m_n = 0.2, and negative-balancing weight λd=250\lambda_d = 250:

    ld(d,d′;s)=λd⋅s⋅max⁡(0,mp−dTd′)+(1−s)⋅max⁡(0,dTd′−mn)l_d(d, d'; s) = \lambda_d \cdot s \cdot \max(0, m_p - d^T d') + (1 - s) \cdot \max(0, d^T d' - m_n)

  4. Knowl 4 — SuperPoint Self-Supervised Training Pipeline

    algorithm

    The SuperPoint training pipeline bootstraps interest point detection on synthetic data, performs self-supervised label generation via Homographic Adaptation on unlabeled real images, and finishes with joint end-to-end detector and descriptor training.

    Input: Synthetic Shapes dataset generator, Unlabeled real image dataset (MS-COCO), Target iterations N_synth, N_real
    Output: Trained SuperPoint joint network f_theta
    // Step 1: Pre-train base detector on synthetic shapes
    Initialize MagicPoint detector f_base (encoder and detector head)
    for step = 1 to N_synth do
        I_synth, Y_synth = GenerateSyntheticShapesWithHomography()
        Update f_base parameters by minimizing L_p(f_base(I_synth), Y_synth)
    end for
    // Step 2: Self-labeling via Homographic Adaptation on target domain
    for each image I in MS-COCO do
        Compute aggregated pseudo-ground-truth heatmap:
        F_hat = (1 / N_h) * sum_{i=1}^{N_h} (H_i^{-1} * f_base(H_i(I)))
        Extract discrete pseudo-ground truth keypoint labels Y_real from F_hat via NMS
    end for
    // Step 3: Iterative refinement (optional)
    Train detector on (MS-COCO, Y_real), update f_base, and re-run Step 2 to generate refined labels Y_real
    // Step 4: Joint training of SuperPoint detector and descriptor
    Initialize SuperPoint network f_theta (shared encoder, detector head, descriptor head)
    for step = 1 to N_real do
        Sample image I from MS-COCO and random pairwise homography H
        I_warp = H(I)
        Y_warp = H(Y_real)
        Compute (X, D) = f_theta(I) and (X_warp, D_warp) = f_theta(I_warp)
        Compute correspondence matrix S from H
        Loss = L_p(X, Y_real) + L_p(X_warp, Y_warp) + lambda * L_d(D, D_warp, S)
        Update f_theta parameters using ADAM optimizer
    end for
    return f_theta

  5. Knowl 5 — Synthetic Shapes Pre-Training and MagicPoint Base Detector

    model/method

    To train an initial interest point detector without human-annotated real data, a synthetic dataset called Synthetic Shapes is created. It consists of synthetically rendered 2D geometric primitives: quadrilaterals, triangles, lines, cubes, checkerboards, stars, ellipses, and noise-only negative images.

    Keypoints are unambiguously defined by geometric junctions (Y-junctions, L-junctions, T-junctions), line segment endpoints, and centers of tiny ellipses. Synthetic images are generated on-the-fly and augmented with random homographic warps so that no sample is seen twice.

    The detector pathway of the SuperPoint architecture (comprising the shared VGG encoder and interest point decoder head, omitting the descriptor head) trained on Synthetic Shapes for 200,000 iterations is named MagicPoint. MagicPoint provides strong corner localization on synthetic primitives and transfers effectively to real images with salient corner structures (e.g., windows, tables), serving as the foundational base detector for self-supervised label generation.

  6. Knowl 6 — HPatches Homography Estimation and Feature Matching Benchmark

    data/table

    The performance of the SuperPoint detector and descriptor network was evaluated on the HPatches benchmark (116 scenes with 696 images) against SIFT, LIFT, and ORB. For each pair of images at 480×640480 \times 640 resolution, up to 1000 keypoints were detected, descriptors matched via nearest neighbors, and homographies estimated using OpenCV's RANSAC-based homography estimator. Correctness is measured at corner error thresholds ϵ∈{1,3,5}\epsilon \in \{1, 3, 5\} pixels, along with Detector Repeatability (Rep.), Mean Localization Error (MLE), Nearest Neighbor mean Average Precision (NN mAP), and Matching Score (M. Score):

    Homography Estimation Detector Metrics Descriptor Metrics
    Method ϵ=1\epsilon=1 ϵ=3\epsilon=3 ϵ=5\epsilon=5 Rep. MLE NN mAP M. Score
    SuperPoint .310 .684 .829 .581 1.158 .821 .470
    LIFT .284 .598 .717 .449 1.102 .664 .315
    SIFT .424 .676 .759 .495 0.833 .694 .313
    ORB .150 .395 .538 .641 1.157 .735 .266

    SuperPoint outperforms LIFT and ORB across all homography correctness thresholds and achieves the highest overall homography accuracy at ϵ=3\epsilon = 3 (.684) and ϵ=5\epsilon = 5 (.829). SIFT performs best at sub-pixel threshold ϵ=1\epsilon = 1 (.424) due to its dedicated sub-pixel localization refinement. SuperPoint achieves substantially higher descriptor quality, scoring .821 in NN mAP and .470 in Matching Score compared to .694 and .313 for SIFT, respectively.

  7. Knowl 7 — Interest Point Detection Repeatability on HPatches Benchmark

    data/table

    Interest point repeatability was measured across the 116 scenes of the HPatches dataset, split into 57 illumination-varying scenes and 59 viewpoint-varying scenes, evaluated at 240×320240 \times 320 resolution with 300 detected points per image and a correctness threshold ϵ=3\epsilon = 3 pixels under Non-Maximum Suppression (NMS) radii of 4 and 8 pixels:

    57 Illumination Scenes 59 Viewpoint Scenes
    Detector NMS=4 NMS=8 NMS=4 NMS=8
    SuperPoint .652 .631 .503 .484
    MagicPoint .575 .507 .322 .260
    FAST .575 .472 .503 .404
    Harris .620 .533 .556 .461
    Shi .606 .511 .552 .453
    Random .101 .103 .100 .104

    SuperPoint achieves the highest repeatability under illumination changes (.652 at NMS=4 and .631 at NMS=8) and demonstrates that Homographic Adaptation substantially improves the base MagicPoint detector, especially on viewpoint changes where repeatability increases from .322 to .503 (NMS=4).

  8. Knowl 8 — Detector Robustness to Noise on Synthetic Shapes Benchmark

    data/table

    The performance of the MagicPoint synthetic base detector was compared against classical corner detectors (FAST, Harris, and Shi-Tomasi) across 10 shape categories (1000 images per category) on Synthetic Shapes under clean and noisy image conditions. Evaluation metrics include mean Average Precision (mAP, higher is better) and Mean Localization Error (MLE in pixels, lower is better) with threshold ϵ=5\epsilon = 5 pixels:

    Metric Noise MagicPointL MagicPointS FAST Harris Shi
    mAP no noise 0.979 0.980 0.405 0.678 0.686
    mAP noise 0.971 0.939 0.061 0.213 0.157
    MLE no noise 0.860 0.922 1.656 1.245 1.188
    MLE noise 1.012 1.078 1.766 1.409 1.383

    MagicPointL (64-128 channel stack) and MagicPointS (9-32 channel stack) substantially outperform classical algorithms. Classical detectors experience catastrophic performance degradation in the presence of noise (e.g., FAST mAP drops from 0.405 to 0.061, Harris from 0.678 to 0.213), whereas MagicPointL remains highly robust, maintaining an mAP of 0.971 under noise.

  9. Knowl 9 — Scale and Sample Size Dynamics in Homographic Adaptation

    empirical result

    Evaluating Homographic Adaptation on a held-out test split of MS-COCO images reveals the following empirical behaviors regarding sample size NhN_h and aggregation type:

    1. Sample Size (NhN_h): Without adaptation (Nh=1N_h = 1), base MagicPoint achieves a repeatability score of 0.67. Increasing the number of random homographies to Nh=100N_h = 100 provides a 21% relative boost in repeatability. Expanding to Nh=1000N_h = 1000 yields a 22% relative boost, demonstrating diminishing returns beyond Nh=100N_h = 100.

    2. Within-Scale vs. Across-Scale Aggregation: Within-scale homographic aggregation operates analogously to set intersection by averaging response maps across geometric transforms at fixed resolution, filtering out non-repeatable spurious responses. Across-scale aggregation behaves analogously to set union by detecting features visible only at specific resolutions in an image pyramid (Ns>1N_s > 1). Experimental isolation demonstrates that within-scale aggregation accounts for the vast majority of repeatability gains.

  10. Knowl 10 — SuperPoint Computational Efficiency and Real-Time Inference

    empirical result

    SuperPoint achieves real-time inference by sharing convolutional encoder computations between detection and description tasks and utilizing non-learned decoders.

    On an NVIDIA Titan X GPU processing 480×640480 \times 640 resolution images:

    • A single forward pass through the shared encoder, interest point decoder, and semi-dense descriptor decoder requires approximately 11.15 ms.
    • Rather than interpolating a dense 480×640×256480 \times 640 \times 256 descriptor volume, sparse sampling of the 1000 detected keypoint locations via CPU bicubic interpolation and L2L_2 normalization takes 1.5 ms.
    • Total latency is approximately 13 ms per image, achieving an overall throughput of ~70 FPS.

Coverage note — Exploratory experiments on synthetic blob detection centers were omitted because they were not included in the main SuperPoint system or evaluation pipeline.

References

  1. 1.V. Badrinarayanan, A. Kendall, and R. Cipolla. SegNet: A deep convolutional encoder-decoder architecture for image segmentation. PAMI, 2017. 3, 8
  2. 2.V. Balntas, K. Lenc, A. Vedaldi, and K. Mikolajczyk. HPatches: A benchmark and evaluation of handcrafted and learned local descriptors. In CVPR, 2017. 7
  3. 3.C. B. Choy, J. Gwak, S. Savarese, and M. Chandraker. Universal Correspondence Network. In NIPS. 2016. 2, 3, 8
  4. 4.D. DeTone, T. Malisiewicz, and A. Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 2
  5. 5.D. DeTone, T. Malisiewicz, and A. Rabinovich. Toward geometric deepslam. arXiv preprint arXiv:1707.07410, 2017. 10
  6. 6.L. F. I. K. P. F. F. M.-N. Edgar Simo-Serra, Eduard Trulls. Discriminative learning of deep convolutional feature point descriptors. In ICCV, 2015. 2
  7. 7.Y. Ganin and V. Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 2
  8. 8.C. Harris and M. Stephens. A combined corner and edge detector. In Alvey vision conference, volume 15, pages 10–5244. Manchester, UK, 1988. 4, 7, 10
  9. 9.R. Hartley and A. Zisserman. Multiple View Geometry in computer vision. 2003. 1
  10. 10.S. Honari, P. Molchanov, S. Tyree, P. Vincent, C. Pal, and J. Kautz. Improving landmark localization with semi-supervised learning. arXiv preprint arXiv:1709.01591, 2017. 2
  11. 11.Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014. 7
  12. 12.C.-Y. Lee, V. Badrinarayanan, T. Malisiewicz, and A. Rabinovich. RoomNet: End-to-end room layout estimation. In ICCV, 2017. 1
  13. 13.T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and L. Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6
  14. 14.W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg. SSD: Single shot multibox detector. In ECCV, 2016. 1, 8
  15. 15.D. G. Lowe. Distinctive image features from scale-invariant keypoints. IJCV, 2004. 2, 7
  16. 16.K. Mikolajczyk and C. Schmid. A performance evaluation of local descriptors. PAMI, 2005. 2, 7, 10
  17. 17.R. Mur-Artal, J. Montiel, and J. D. Tardos. ORB-SLAM: a versatile and accurate monocular SLAM system. IEEE Transactions on Robotics, 2015. 7
  18. 18.A. Odena, V. Dumoulin, and C. Olah. Deconvolution and checkerboard artifacts. Distill, 2016. 3
  19. 19.A. Paszke, S. Gross, S. Chintala, and G. Chanan. PyTorch. https://github.com/pytorch/pytorch. 6
  20. 20.I. Rocco, R. Arandjelovic, and J. Sivic. Convolutional neural network architecture for geometric matching. In CVPR, 2017. 2
  21. 21.E. Rosten and T. Drummond. Machine learning for high-speed corner detection. In ECCV, 2006. 2, 4, 7, 10
  22. 22.E. Rublee, V. Rabaud, K. Konolige, and G. Bradski. ORB: An efficient alternative to SIFT or SURF. In ICCV, 2011. 7
  23. 23.N. Savinov, A. Seki, L. Ladicky, T. Sattler, and M. Pollefeys. Quad-networks: unsupervised learning to rank for interest point detection. In CVPR. 2017. 2
  24. 24.C. Schmid, R. Mohr, and C. Bauckhage. Evaluation of interest point detectors. IJCV, 2000. 2
  25. 25.J. Shi and C. Tomasi. Good features to track. In CVPR, 1994. 4, 7, 10
  26. 26.W. Shi, J. Caballero, F. Huszár, J. Totz, A. P. Aitken, R. Bishop, D. Rueckert, and Z. Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 3
  27. 27.K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3, 6
  28. 28.B. Ummenhofer, H. Zhou, J. Uhrig, N. Mayer, E. Ilg, A. Dosovitskiy, and T. Brox. DeMoN: Depth and motion network for learning monocular stereo. In CVPR, 2017. 2
  29. 29.Y. Verdie, K. Yi, P. Fua, and V. Lepetit. TILDE: A Temporally Invariant Learned DEtector. In CVPR, 2015. 2
  30. 30.S. Vijayanarasimhan, S. Ricco, C. Schmid, R. Sukthankar, and K. Fragkiadaki. SfM-Net: Learning of structure and motion from video. arXiv preprint arXiv:1704.07804, 2017. 2
  31. 31.S.-E. Wei, V. Ramakrishna, T. Kanade, and Y. Sheikh. Convolutional pose machines. In CVPR, 2016. 1
  32. 32.K. M. Yi, E. Trulls, V. Lepetit, and P. Fua. LIFT: Learned Invariant Feature Transform. In ECCV, 2016. 2, 7, 8
  33. 33.T. Zhou, M. Brown, N. Snavely, and D. G. Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 2

Citation

MLA
DeTone, D., et al. “SuperPoint: Self-Supervised Interest Point Detection and Description”. arXiv, 2017, http://arxiv.org/abs/1712.07629v4.
APA
DeTone, D., Malisiewicz, T., & Rabinovich, A. (2017). SuperPoint: Self-Supervised Interest Point Detection and Description. arXiv. http://arxiv.org/abs/1712.07629v4
Chicago
DeTone, D., T. Malisiewicz, and A. Rabinovich. 2017. “SuperPoint: Self-Supervised Interest Point Detection and Description”. arXiv. http://arxiv.org/abs/1712.07629v4.
Harvard
DeTone, D., Malisiewicz, T. and Rabinovich, A. (2017) “SuperPoint: Self-Supervised Interest Point Detection and Description”, arXiv [Preprint]. Available at: http://arxiv.org/abs/1712.07629v4.
Vancouver
1. DeTone D, Malisiewicz T, Rabinovich A (2017) SuperPoint: Self-Supervised Interest Point Detection and Description. arXiv

BibTeX

@article{detone2017superpoint,
  title = {SuperPoint: Self-Supervised Interest Point Detection and Description},
  author = {DeTone, Daniel and Malisiewicz, Tomasz and Rabinovich, Andrew},
  year = {2017},
  journal = {arXiv},
  url = {http://arxiv.org/abs/1712.07629v4},
  eprint = {1712.07629}
}
Metadata:arXiv

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: IEEE