Scale-Invariant Detection

Scale-Invariant Detection

Scale-invariant detection finds image features whose response is stable across zoom — a keypoint detected at full resolution reappears at a correspondingly smaller location when the image is downsampled. Harris corners are rotation-invariant but not scale-invariant: zooming in on a corner eventually turns it into an edge, destroying the response.

The standard scale-invariant detector operates on a Gaussian scale-space:
$L(x,y,\sigma )=G(x,y,\sigma )\ast I(x,y),$
and searches for extrema of the scale-normalized Laplacian ${\sigma }^2{\nabla }^{2}L$ jointly over spatial and scale dimensions. Local maxima in $(x,y,\sigma )$-space identify blob-like features at their characteristic scales.

In practice, the Laplacian is approximated by the Difference of Gaussians (DoG):
$D(x,y,\sigma )=L(x,y,k\sigma )-L(x,y,\sigma )\approx (k-1){\sigma }^2{\nabla }^{2}L,$
which is cheap to compute as a pixel-wise subtraction between blurs. Extrema of $D$ over the $3\times 3\times 3$ spatial-scale neighborhood are the keypoints; this is the detection stage of SIFT.

Intuition

The same feature at different zoom levels is represented by different spatial extents: a circular blob 10 pixels across at one zoom becomes 20 pixels across when you zoom in 2×. A filter matched to the smaller blob will miss the larger one and vice versa. Scale-invariant detection solves this by sweeping across a family of filter sizes (or, equivalently, a family of blurs) and recording where each feature is most prominent.

The Laplacian is the right filter for this sweep: it responds strongly when its scale matches the blob’s scale, and weakly otherwise. Tracking the response across scales and recording the scale at which it peaks gives both the feature’s location and its characteristic scale — the latter tells the descriptor stage how large a neighborhood to sample.

The DoG approximation is what makes this practical. Computing the true scale-normalized Laplacian at many scales requires convolving with second-derivative kernels; the DoG is just blur-differencing, an operation the Gaussian pyramid already produces as a side effect.

Properties

Scale Space

(Gaussian pyramid) Successive blurs with scale factor $k$ produce a discrete sampling of scale space.

Typical parameters: ${\sigma }_0=1.6$, $s=3$ scales per octave, $k=2^{1/s}$. An octave is a doubling of scale; after each octave the image is downsampled by 2 to save compute.

(Characteristic scale) The scale at which a feature’s normalized Laplacian response peaks is its intrinsic size.

Lindeberg’s scale-space theory shows that this characteristic scale covaries with image zoom: halving the image resolution doubles each feature’s characteristic scale. This covariance is what makes the detector scale-invariant.

DoG Extrema

(Spatial-scale search) A pixel is a keypoint if its DoG value is a local extremum over the $3\times 3\times 3$ neighborhood across spatial and scale dimensions.

The middle scale’s $3\times 3$ spatial neighborhood plus the $9$ pixels in each of the adjacent scales gives $26$ comparisons per candidate. Implementing this as a single sweep keeps it efficient.

(Subpixel / subscale refinement) A Taylor expansion of $D$ around the discrete extremum locates the continuous maximum.

The refined $(x,y,\sigma )$ gives keypoints whose location is accurate to a fraction of a pixel and a fraction of an octave — important for matching across very different zoom levels.

(Contrast and edge rejection) Low-contrast extrema and edge-like responses are filtered out.

Low-contrast points are unstable under noise; edge extrema have one small and one large principal curvature and can be flagged by the Hessian trace-determinant ratio used in Harris.

Alternatives

(Hessian-Laplace, Harris-Laplace) Combine a corner detector with a scale-selection step.

Harris-Laplace runs Harris at each scale and keeps responses that are simultaneously Harris-maximal in space and Laplacian-maximal in scale. Produces fewer but more precise keypoints than DoG.

(MSER — Maximally Stable Extremal Regions) Detects blob regions whose intensity thresholding is stable across a range of thresholds.

Complementary to DoG: MSER excels on text, signs, and high-contrast blobs; DoG is more general-purpose.

(Learned detectors) SuperPoint and similar CNNs learn scale-invariant keypoint heatmaps jointly with descriptors.

Typically outperform DoG on hard benchmarks but require GPU inference.

See Also

• 4.1.1. Harris Corner Detection — the rotation-invariant detector this extends with scale
• 4.3.2. SIFT — where DoG-detected keypoints are paired with a rich descriptor
• 4.1. Feature Detection — the broader family of detection strategies