BRISK

BRISK
Name	BRISK
Type	Computer vision, Feature detection, Feature description
Introduced	2011
Authors	Stefan Leutenegger, Margarita Chli, Roland Y. Siegwart
Publication	IEEE International Conference on Computer Vision (ICCV) 2011
Influenced by	FAST, SIFT, SURF
Programming languages	C++, Python, MATLAB
License	BSD-style (varies by implementation)

BRISK

BRISK is a local feature detector and binary descriptor designed for real-time image matching and robust feature description. Developed by Stefan Leutenegger, Margarita Chli, and Roland Y. Siegwart and presented at IEEE International Conference on Computer Vision in 2011, BRISK aims to combine fast keypoint detection with compact, rotation- and scale-aware binary descriptors for use in applications such as visual odometry, simultaneous localization and mapping, and image retrieval. It builds upon prior work such as FAST (corner detector), SIFT (algorithm), and SURF (speeded up robust features), seeking a trade-off between computational efficiency and matching robustness.

Introduction

BRISK was introduced to address the need for efficient local features in robotic perception tasks exemplified by development in ETH Zurich labs and robotics conferences like ICRA and RSS. The method integrates a scale-space adaptation similar to approaches in Lowe's SIFT scale-selection while leveraging a pattern of predetermined sampling points inspired by descriptor ideas in BRIEF (binary robust independent elementary features) and binary descriptors such as ORB (oriented FAST and rotated BRIEF). BRISK is often discussed alongside other binary descriptors like FREAK (fast retina keypoint) and BRIEF (binary robust independent elementary features) in literature surveys and benchmarks such as those at CVPR workshops and datasets from Oxford Visual Geometry Group.

Algorithm

The BRISK pipeline consists of two main components: keypoint detection and descriptor computation. For detection, BRISK adapts the corner-response strategy from FAST (corner detector) to a scale-space pyramid influenced by multiscale concepts in SIFT (algorithm) and Harris corner detector, performing non-maximum suppression across scales and image positions similar to methods used in DoG (difference of Gaussians)-based detectors. Keypoint orientation estimation exploits local gradient approximations comparable to orientation assignment in SIFT (algorithm) and moment-based schemes used in Harris corner detector research.

Descriptor computation samples a predefined concentric pattern of smoothed intensity pairs around each keypoint, deriving binary tests via pairwise intensity comparisons. This sampling scheme echoes retina-inspired sampling in FREAK (fast retina keypoint) and the rotational invariance strategies in ORB (oriented FAST and rotated BRIEF). The descriptor encodes tests into a compact binary string enabling Hamming-distance matching as used in place-recognition systems like those evaluated on KITTI (dataset) and TUM RGB-D benchmarks. BRISK also includes mechanisms for scale normalization by selecting an appropriate octave and pattern scaling analogous to procedures in SIFT (algorithm) and SURF (speeded up robust features).

Variants and Extensions

Several variants extend BRISK for improved robustness, speed, or specificity. Rotated and scale-invariant modifications draw on orientation strategies from ORB (oriented FAST and rotated BRIEF) and scale-space ideas from SIFT (algorithm). Hybrid descriptors combine BRISK keypoints with descriptors like FREAK (fast retina keypoint), BRIEF (binary robust independent elementary features), or floating-point descriptors such as SIFT (algorithm)-like histograms. GPU-accelerated implementations port BRISK computations to frameworks such as CUDA and OpenCL, similar to parallelized efforts for SURF (speeded up robust features) and ORB (oriented FAST and rotated BRIEF). Extensions for illumination invariance and robustness to blur take inspiration from preprocessing ideas in Retina-inspired models and adaptive smoothing techniques used in Hessian Affine detectors.

Performance and Evaluation

Benchmark evaluations compare BRISK against detectors and descriptors including SIFT (algorithm), SURF (speeded up robust features), ORB (oriented FAST and rotated BRIEF), FREAK (fast retina keypoint), and BRIEF (binary robust independent elementary features). On datasets such as those from Oxford Visual Geometry Group, KITTI (dataset), and TUM RGB-D, BRISK typically achieves favorable trade-offs: lower computational cost and compact descriptors versus slightly reduced distinctiveness relative to floating-point descriptors like SIFT (algorithm). In robotic visual odometry pipelines exemplified by systems from ETH Zurich and research groups at MPI for Robotics, BRISK's binary nature yields fast Hamming matching suitable for real-time constraints and embedded platforms such as those using ARM processors and low-power compute modules showcased at ISMAR and IROS.

Applications

BRISK has been applied in diverse domains: visual odometry and SLAM in research from ETH Zurich and institutes using datasets like KITTI (dataset) and TUM RGB-D; augmented reality workflows demonstrated at ISMAR and SIGGRAPH-adjacent venues; image retrieval experiments in university labs that reference Oxford Visual Geometry Group benchmarks; and mobile vision applications on devices from manufacturers featured at Mobile World Congress. Robotics projects at institutions such as Carnegie Mellon University, Stanford University, and ETH Zurich have integrated BRISK for real-time tracking, mapping, and loop-closure detection alongside descriptors like ORB (oriented FAST and rotated BRIEF) and FREAK (fast retina keypoint).

Implementation and Availability

Open-source implementations of BRISK are available in major computer vision libraries including OpenCV, with contributions from communities that include developers associated with Google Summer of Code and research code releases from authors hosted on platforms like GitHub. Implementations exist in C++, Python, and MATLAB, and GPU-accelerated variants leverage CUDA or OpenCL. Licensing is typically permissive (BSD-style) in popular distributions, enabling use in academic projects at institutions such as ETH Zurich, CMU, and Stanford University and in industry prototypes showcased at conferences like CVPR and ICCV.

Category:Computer vision