AI Glossary / Computer vision and multimodal AI / Advanced
Birchfield - Tomasi dissimilarity
Definition
Birchfield - Tomasi dissimilarity is a local image-matching cost designed to reduce sensitivity to how a continuous image signal is sampled onto pixel grids. It compares pixel intensities with nearby linearly interpolated intensity values. It is associated particularly with stereo matching, where corresponding scene points may not align exactly with pixel centers.
What it means
Two cameras can observe the same surface feature yet record different center-pixel intensities because the feature falls at slightly different subpixel positions. A plain absolute intensity difference can penalize that sampling difference as though it were evidence of a bad match. Birchfield - Tomasi comparison allows a small local interpolation range before assigning the cost. It addresses a specific matching problem; it is not a complete depth algorithm or a general guarantee against lighting changes and image noise.
How it works
Along corresponding rectified scanlines, form the range of linearly interpolated intensities within half a pixel of each candidate location. Compare the left center intensity with the right range, and the right center intensity with the left range. Each directional distance is zero when the center value falls inside the other range, otherwise its distance to the nearest endpoint. The symmetric cost takes the smaller directional distance. A stereo method then uses such costs alongside its search, aggregation, regularization and validity checks. A low local cost alone does not establish a unique correspondence.
Three practical examples
- A stereo developer compares absolute intensity difference with Birchfield - Tomasi costs on image pairs containing fractional-pixel shifts, keeping the rest of the pipeline fixed.
- An engineer investigates why a depth algorithm produces mismatches near a sharp brightness transition and checks whether image sampling contributes to the problem.
- A researcher uses an established stereo implementation whose documented matching cost includes Birchfield - Tomasi comparison, while distinguishing that local component from the complete algorithm's behavior.
Think of it this way
Think of comparing two readings taken on slightly shifted rulers. Looking only at the nearest marked tick can exaggerate disagreement. Allowing a small interpolation between neighboring readings can reveal that the underlying signals are compatible. This does not prove that the rulers refer to the same object; it simply makes one part of the comparison less sensitive to where the marks happen to fall.
Worked example
In an original grayscale example, a right-image candidate has neighboring intensities 80, 100 and 140, with 100 at the center. Halfway interpolations are 90 and 120, so its local range is [90,120]. Let the left center intensity be 110. Its distance to that range is zero, although the center-to-center absolute difference is |110−100|=10. Therefore the final symmetric cost is zero regardless of the nonnegative reverse directional distance. This illustrates sampling tolerance. A zero cost does not establish that the candidate is the true match: another location might have a compatible range too, and the full stereo process must resolve ambiguity.
Illustrative case study
A fictional robotics team evaluates a classical stereo pipeline on a controlled tabletop scene. It notices that some intensity-based costs change sharply when the cameras shift by a small amount. The team tests Birchfield - Tomasi comparison as one component while preserving calibration, disparity range and evaluation data. It measures the final disparity errors against references and examines occlusions separately. The report avoids claiming that any improvement comes from solving all stereo failure modes. This case illustrates a component-level experiment, not a reported benchmark result or a promise of better navigation performance.
Common misconceptions and mistakes
The measure does not replace camera calibration or image rectification. It also does not solve occlusion, repeated patterns or missing texture by itself. Calling it sampling-insensitive should not be read as unconditional invariance to every image transformation. Another error is confusing the local cost with the final disparity estimate. A pipeline may transform images, aggregate costs and impose smoothness before selecting disparities, so its overall behavior cannot be explained by this one formula alone. Implementation details should be checked against the actual library version used.
Business use
This term is most useful when evaluating or maintaining a stereo-vision system, rather than as a standalone product capability. Understanding the cost helps technical teams distinguish a sampling issue from calibration, lighting or search problems. A business-facing evaluation should focus on final depth reliability, runtime and failure handling under the intended conditions. If a supplier names Birchfield - Tomasi dissimilarity, that identifies one algorithmic ingredient; it does not establish the complete system's accuracy or suitability for a measurement task.
People also ask
Is the output a depth value?
No. It is a dissimilarity cost for candidate image matches.
Does zero mean the pixels must correspond?
No. It means the local comparison found compatible intensity values under its rule.
Is it used in practical software?
OpenCV documents a Birchfield - Tomasi cost within its modified StereoSGBM implementation, alongside other processing stages.
Check your understanding
Why was the worked directional cost zero?
Answer: The left intensity 110 lay inside the right interpolation range [90,120].
Why was absolute difference still ten?
Answer: It compared only the two center samples.
What remains after computing a local cost?
Answer: Selecting consistent correspondences and handling ambiguity, occlusion and other errors in the full stereo pipeline.
Related terms
Learn next
Calculate center-pixel absolute differences and interval distances for a few short scanlines. Shift a synthetic brightness edge by a fraction of a pixel and compare the costs. Then trace how a complete stereo implementation turns local costs into a disparity map.
More in Computer vision and multimodal AI
- Binary image - A binary image is an image whose pixels have only two possible values.
- Binocular disparity - Binocular disparity is the difference between the positions of corresponding scene features in two views taken from separated viewpoints.
- Binocular stereopsis - Binocular stereopsis is the perception of depth from differences between the images received by the two eyes.
- Blind deconvolution - Blind deconvolution is the estimation of a sharper underlying signal or image when the blurring process is also unknown or only partly known.
- Block matching - Block matching is an image-processing method that compares a small rectangular patch with candidate patches to find similar content.
- Block-matching and 3-D filtering (BM3D) - Block-matching and 3-D filtering, usually called BM3D, is an image-denoising method that groups similar two-dimensional patches into three-dimensional...
Sources
Birchfield and Tomasi, A Pixel Dissimilarity Measure That Is Insensitive to Image Sampling: https://ai.stanford.edu/~birch/publications/dissimilarity_pami1998.pdf OpenCV, StereoSGBM class documentation: https://docs.opencv.org/4.10.0/d2/d85/classcv_1_1StereoSGBM.html The intensity values, calculation and robotics scenario are original educational illustrations. Software details refer to the cited documentation rather than a universal implementation.
Previous: Bipartite Next: Birnbaum's theorem