CoChat

SURF: Speeded Up Robust Features

Open access
Bay, Herbert, Tuytelaars, Tinne, Van Gool, Luc · 2006 · Lecture Notes in Computer Science

The DOI matches its Crossref record; 43 of the 43 references it cites were found in a registry or an index; none of the 18 sources its claims cite is checked for support yet.

Overview

CoChat Summary

TL;DR

SURF is a scale- and rotation-invariant interest point detector and descriptor that uses integral images, Hessian-based detection, and Haar-wavelet responses to be faster than prior methods while matching or exceeding their repeatability and distinctiveness. It reports strong results on standard benchmarks and a museum object recognition task.

Key findings

  • The Fast-Hessian detector is more than 3× faster than DoG and 5× faster than Hessian-Laplace, with comparable or better repeatability on Graffiti, Leuven, Boat, and Wall; on the first Graffiti image (800×640), computation times were 120 ms for Fast-Hessian, 400 ms for DoG, 650 ms for Hessian-Laplace, and 1800 ms for Harris-Laplace.
  • SURF’s 64-dimensional descriptor outperformed GLOH, SIFT, and PCA-SIFT in most comparisons, sometimes by more than 10% improvement in recall for the same precision; SURF-128 was slightly better but slower to match.
  • Joint detector-descriptor computation times on the first Graffiti image were 255 ms for U-SURF, 354 ms for SURF, 391 ms for SURF-128, and 1036 ms for SIFT.
  • In museum object recognition, recognition rates were 85.7% for SURF-128, 83.8% for U-SURF, 82.6% for SURF, 78.3% for GLOH, 78.1% for SIFT, and 72.3% for PCA-SIFT.

Methods

  • Design: a scale- and rotation-invariant detector/descriptor. The detector uses the determinant of the Hessian with box-filter approximations evaluated via integral images; the descriptor uses Haar-wavelet responses in a 4×4 subregion grid, producing a 64-dimensional vector, with an upright U-SURF variant and an extended SURF-128 variant.
  • Data: standard evaluation sequences from Mikolajczyk (including Graffiti, Wall, Boat, Leuven, Bikes, Trees, and Ubc; Bark excluded from descriptor evaluations) and a real-life museum object recognition database of 216 images of 22 objects, with 116 test images at 320×240.
  • Analysis: detector repeatability scores, descriptor recall-(1-precision) graphs under similarity-threshold and nearest-neighbour-ratio matching, recognition rates, and computation timings on a standard Linux PC (Pentium IV, 3 GHz).

Limitations

  • The paper has no explicit limitations section. It states that space limitations prevent showing results on all sequences, and it assumes skew, anisotropic scaling, and perspective effects are second-order effects covered only by overall descriptor robustness.
  • Evident from the text: the method is scale- and rotation-invariant but not affine-invariant; Fig. 4 notes the interest points are not affine invariant, and the descriptor does not use colour.
  • Evident from the text: the object recognition experiment uses small 320×240 test images and matching without additional geometric constraints, which the authors say would reduce false positives but were omitted for comparison purposes.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1007/11744023_32

Cites

43 of 43 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Left to right: The (discretised and cropped) Gaussian second order partial derivatives in y -direction and xy -direction, and our approximations thereof using box filters. The grey regions are equal to zero.
Fig. 1 Left to right: The (discretised and cropped) Gaussian second order partial derivatives in y -direction and xy -direction, and our approximations thereof using box filters. The grey regions are equal to zero.From the PDF
Image 2
From the PDF
Fig. 2 Left: Detected interest points for a Sunflower field. This kind of scenes shows clearly the nature of the features from Hessian-based detectors. Middle: Haar wavelet types used for SURF. Right: Detail of the Graffiti scene showing the size of the descriptor window at different scales.
Fig. 2 Left: Detected interest points for a Sunflower field. This kind of scenes shows clearly the nature of the features from Hessian-based detectors. Middle: Haar wavelet types used for SURF. Right: Detail of the Graffiti scene showing the size of the descriptor window at different scales.From the PDF
Fig. 3 The descriptor entries of a sub-region represent the nature of the underlying intensity pattern. Left: In case of a homogeneous region, all values are relatively low. Middle: In presence of frequencies in x direction, the value of ∑ | d x | is high, but all others remain low. If the intensity is gradually increasing in x direction, both values ∑ d x and ∑ | d x | are high.
Fig. 3 The descriptor entries of a sub-region represent the nature of the underlying intensity pattern. Left: In case of a homogeneous region, all values are relatively low. Middle: In presence of frequencies in x direction, the value of ∑ | d x | is high, but all others remain low. If the intensity is gradually increasing in x direction, both values ∑ d x and ∑ | d x | are high.From the PDF
Image 5
From the PDF
Fig. 5 An example image from the reference set (left) and the test set (right). Note the difference in viewpoint and colours.
Fig. 5 An example image from the reference set (left) and the test set (right). Note the difference in viewpoint and colours.From the PDF
Fig. 6 Repeatability score for image sequences, from left to right and top to bottom, Wall and Graffiti (Viewpoint Change), Leuven (Lighting Change) and Boat (Zoom and Rotation)
Fig. 6 Repeatability score for image sequences, from left to right and top to bottom, Wall and Graffiti (Viewpoint Change), Leuven (Lighting Change) and Boat (Zoom and Rotation)From the PDF
Fig. 7 Recall, 1-Precision graphs for, from left to right and top to bottom, Viewpoint change of 50 (Wall) degrees, scale factor 2 (Boat), image blur (Bikes and Trees), brightness change (Leuven) and JPEG compression (Ubc)
Fig. 7 Recall, 1-Precision graphs for, from left to right and top to bottom, Viewpoint change of 50 (Wall) degrees, scale factor 2 (Boat), image blur (Bikes and Trees), brightness change (Leuven) and JPEG compression (Ubc)From the PDF

Verification

Claims the paper makes

0 Valid · 2 Not valid · 12 Not checked · 30 Own results and interpretations

8 more claims

Create a free account to see all 44 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 24
  • Probably not a substantial claim. 13
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Citations

Cites

Show all 43 references

Cited by

Items 12,501–12,600 of 14,752

Metadata

Authors
Bay, H., Tuytelaars, T., & Van Gool, L.
Affiliations
  • ETH Zurich (CH)
  • KU Leuven (BE)
Venue
Lecture Notes in Computer Science
Published
2006
Pages
404-417
ISSN
  • 0302-9743 (ISSN-L)
  • 1611-3349
OpenAlex
W1677409904
Cited by
14,955
References
39
Citations by year
2026: 210 · 2025: 519 · 2024: 727 · 2023: 757 · 2022: 878 · 2021: 975 · 2020: 978 · 2019: 1036 · 2018: 968 · 2017: 976 · 2016: 930 · 2015: 1079 · 2014: 1006 · 2013: 982 · 2012: 884
FWCI
196.80
Type
conference paper
Language
en
Open access
Licence
www.springer.com/tdm (tdm)
Locations
  • Lecture notes in computer science (publishedVersion)
  • Lirias (KU Leuven) (submittedVersion, other-oa)
Funders
Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Image and Video Retrieval Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image Retrieval and Classification Techniques
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Image and Object Detection Techniques
Keywords
SURF, keypoint detection, feature descriptors, scale invariance, rotation invariance, integral image, Hessian matrix, feature matching, object recognition, repeatability

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.