MOPS

MOPS

Multi-Scale Oriented Patches (MOPS) is a simple rotation- and illumination-invariant feature descriptor built on Harris corners. Each descriptor is a bias-and-gain-normalized $8\times 8$ image patch, sampled from a larger oriented window around the keypoint.

For each detected keypoint:

1. Compute orientation. Estimate the dominant gradient direction in a small window around the keypoint; this becomes the patch’s canonical orientation.
2. Extract an oriented window. Sample a $40\times 40$ pixel window rotated to the canonical orientation and centered on the keypoint.
3. Downsample. Gaussian-blur and subsample the window to $8\times 8$ pixels. The blur-then-subsample step acts as a low-pass anti-aliasing filter.
4. Normalize. Subtract the patch mean and divide by its standard deviation:
   $\tilde{I}=(I-\mu )/\sigma ,$
   making the descriptor invariant to affine intensity changes.

The resulting 64-dimensional vector is matched across images by Euclidean distance, typically filtered with Lowe’s ratio test.

Intuition

MOPS is the lightweight cousin of SIFT: same pipeline shape (orient, sample, normalize, describe) but replacing the gradient-histogram descriptor with a raw normalized patch. The result is a much faster descriptor with a coarser invariance profile — it handles rotation and linear illumination change but is more brittle than SIFT to small spatial deformations, scale change outside the sampling octave, and nonlinear lighting.

The key design choice is the large-window-then-downsample pattern. Sampling a $40\times 40$ window and reducing it to $8\times 8$ averages over small alignment errors and gives the descriptor graceful tolerance to subpixel misregistration of the keypoint itself.

Properties

(Orientation assignment) The dominant gradient in a $3\times 3$ or $5\times 5$ window defines the patch’s canonical axis.

The sampled window is then rotated so this direction is “up” before downsampling, giving rotation invariance. Unlike SIFT, MOPS does not form an orientation histogram — it uses a single gradient direction.

(Anti-aliasing via Gaussian) The $40\to 8$ downsampling is preceded by Gaussian blur to prevent aliasing.

Without blurring, direct subsampling of a $40\times 40$ window aliases high-frequency detail into the 8 × 8 descriptor, producing noisy features that match poorly across images.

(Intensity normalization) Mean-subtraction and standard-deviation division give invariance to affine illumination change $I\to aI+b$.

This handles exposure differences, gain control, and uniform brightness shifts between images but does not compensate for nonlinear lighting (highlights, shadows, vignetting).

(Simpler and faster than SIFT) A 64-D patch descriptor is cheaper than a 128-D gradient histogram but less distinctive.

MOPS is commonly used in teaching pipelines and in applications where compute is tight and scenes are well-lit (e.g., panorama stitching of travel photos). For demanding matching problems, SIFT or a learned descriptor is preferred.

See Also

• 4.1.1. Harris Corner Detection — the detector most commonly paired with MOPS
• 4.3.2. SIFT — the richer descriptor that superseded MOPS in most pipelines
• 4.3. Feature Description — context within the broader descriptor family