Modern features: advances, applications, and software


Organizers: Andrea Vedaldi (Univ. of Oxford, UK), Jiri Matas, Krystian Mikolajczyk, Tinne Tuytelaars, Cordelia Schmid, Andrew Zisserman (Univ. of Oxford, UK)
Duration: half day
Abstract: This course will introduce local feature detectors and descriptors as foundational tools in a variety of state-of-the-art computer vision applications. The first part of the tutorial will cover popular co-variant detectors (Harris, Laplacian, Hessian corners and blobs, scale and affine adaptation, MSER, SURF, FAST, etc.) and descriptors (SIFT, SURF, BRIEF, LIOP, etc.), with a particular emphasis on recent advances and additions to this set of tools. It will be shown how the various methods achieve different trade-offs in repeatability, speed, geometric accuracy, and applicability to different image contents in term of their performance in benchmarks and applications (tracking, reconstruction, retrieval, stitching, text detection in the wild, etc.). The second part of the tutorial will review software for computing local features and evaluating their performance automatically on benchmark data. In particular, two software resources will be introduced to the community for the first time: a novel extension to the popular open-source VLFeat library containing new reference implementations of co-variant feature detectors; and a novel benchmarking software superseding standard packages for the evaluation of co-variant feature detectors and descriptors.


  • Part I: modern local features and applications [2 hours]
    • Local features: why, what, and how well
      • Covariant feature frames
      • Invariant feature descriptors
      • Applications
        • tracking, reconstruction, retrieval, stitching, text detection in the wild, …
      • Goals of good feature detectors and descriptors
        • density/repeatability trade-off, speed, geometric accuracy
      • Evaluation
        • ad-hoc benchmarks
        • test applications
    • Co-variant feature detectors
      • Corners and blobs: Harris, Laplacian, Hessian detectors
      • Scale adaptation: Laplacian, Hessian
      • Affine adaptation: structure tensor
      • MSER
      • diffeomorphic co-variants
      • The speed-accuracy: SURF, FAST
    • Local feature descriptors
  • Part 2: software [1 hour]
    • VLFeat and recent extensions
      • A new reference implementation of affine co-variants
      • Descriptors (SIFT, BRIEF, LIOP)
    • Other useful components
      • Quantization and vocabularies
        • k-means
        • randomized kd-trees
      • Performance evaluation
        • PR and ROC curves
    • Demonstrations: a large scale indexing system
      • Oxford 5k dataset
      • feature extraction
      • inverted index
      • spatial verification
      • evaluation

Course website:


  • Refer to the course website for all the materials