AI Glossary / Computer vision and multimodal AI / Intermediate
Condensation algorithm
Definition
The Condensation algorithm is a probabilistic visual-tracking method that represents uncertainty with a collection of weighted samples. Its name abbreviates conditional density propagation. It carries several possible object states forward through time and updates their support using new image evidence, rather than committing immediately to one interpretation.
What it means
In a cluttered image, several shapes or locations may resemble the target. A tracker that keeps only one estimate can follow the wrong object and lose the correct alternative.
Condensation represents a probability distribution using samples, often called particles. Each sample encodes a possible state, such as the position and shape parameters of a tracked contour. Weights indicate how well those possibilities agree with the available evidence under the model.
The method is an influential application of particle filtering to vision. It is not a process for compressing a dataset, despite the everyday meaning of “condensation.”
How it works
The tracker starts with a distribution over plausible states. At each step, it selects previous samples according to their weights, propagates them through a motion model with uncertainty, and evaluates their agreement with the new image.
Samples that better explain the observation receive greater weight. Repeating this process approximates the changing conditional distribution of the target state given the image sequence.
Unlike a single Gaussian description, a sample collection can represent several separated hypotheses. However, a finite sample set can miss important possibilities, especially when the state is high-dimensional or the motion model predicts the wrong region.
The observation model and dynamics are therefore central. More particles can improve approximation in some situations, but they cannot automatically repair a poor model or completely uninformative observations.
Three practical examples
1. Contour tracking: A research system maintains alternative positions of an object's outline when background edges create ambiguous matches.
2. Motion analysis: A laboratory tracks a moving marker while representing uncertainty during brief periods of visual clutter.
3. Interactive prototypes: Designers test whether a probabilistic tracker can maintain a gesture-related state when several nearby image structures look similar.
Think of it this way
Think of keeping several plausible routes on a map while following incomplete clues. New evidence gives some routes more support, while others become less likely.
The analogy is limited because particles are numerical samples, not conscious guesses. Their survival depends on probability models and sampling. A plausible route can disappear from a finite sample set even when it later turns out to be correct.
Worked example
Suppose four equally weighted predicted particles represent horizontal positions 10, 12, 14, and 16 pixels. A new image gives them likelihood values 0.1, 0.6, 0.2, and 0.1.
Because the prior weights are equal and these likelihoods sum to one, the normalized updated weights are also 0.1, 0.6, 0.2, and 0.1. Their weighted mean position is 10 × 0.1 + 12 × 0.6 + 14 × 0.2 + 16 × 0.1 = 12.6 pixels.
If four particles are resampled, the expected copy counts are 0.4, 2.4, 0.8, and 0.4, but actual counts are whole numbers and depend on the sampling procedure.
The next motion step spreads or moves these samples. This simplified calculation illustrates weighting and resampling; a real contour state has more dimensions and a richer image-likelihood model. A mean can also be misleading when separate hypotheses remain.
Illustrative case study
Fictional case: Acorn Vision tests a tracker on a moving shape against a patterned background. A single-estimate baseline sometimes jumps to a similar background edge.
The team evaluates a particle-based tracker that retains alternative contour positions. During a rapid motion, however, all particles remain near the old location and the target is lost.
Researchers examine the motion model, particle coverage, and a separate reinitialization strategy. They measure recovery and uncertainty rather than showing only favorable video segments. The fictional case illustrates why multiple hypotheses help but do not guarantee robust tracking in every sequence.
Common misconceptions and mistakes
The highest-weight particle is not proof of the true state. It is the strongest represented hypothesis under the current model and evidence.
Averaging particles does not always produce a meaningful state. Two plausible locations can have an average between them where no target exists.
More particles do not compensate for incorrect assumptions indefinitely, and they increase computational cost.
Do not confuse uncertainty represented by particles with a calibrated confidence guarantee. Calibration and tracking quality need evaluation against appropriate ground truth.
Business use
The practical lesson is valuable beyond this historical algorithm: visual tracking should represent ambiguity and account for time. A business evaluating a tracker should test clutter, rapid movement, partial occlusion, and recovery after loss.
For a Condensation-style implementation, assess the state representation, observation model, dynamics, sample budget, and initialization. Report latency and failure behavior as well as average position error. Provide a way to declare uncertainty or loss rather than always returning a confident-looking location.
People also ask
Is Condensation related to particle filters?
Yes. It uses sample-based probability propagation for visual tracking and is commonly understood in that family.
Can it keep multiple possible locations?
Yes, if the samples adequately represent them. Resampling can still eliminate weakly supported alternatives.
Does it recognize an object's identity by itself?
Not necessarily. Tracking a modeled state and recognizing identity are different tasks that may require separate components.
Check your understanding
Two distant particle clusters have similar weight. Why might their weighted mean be a poor location estimate?
Answer: The mean can fall between the clusters, where there is little or no support for the target. The distribution's multiple hypotheses should be considered rather than summarized blindly.
Related terms
- Bayesian filtering
- Visual tracking
- Resampling
- Motion model
- Particle filter
- importance weight
- observation likelihood
- multimodal distribution
Learn next
Learn Bayes' rule and then trace one predict - weight - resample cycle with a few particles. Compare a single Gaussian estimate with two separated particle clusters. Examine how motion uncertainty and the number of particles affect the ability to retain alternative tracking hypotheses.
More in Computer vision and multimodal AI
- Computational photography - Computational photography uses algorithms together with captured light measurements to form, improve, or extend photographic images.
- Computer vision - Computer vision is the field of developing methods that extract useful information from images, video and other visual measurements.
- Computer vision dazzle - Computer vision dazzle is a visual camouflage approach that explores whether changes in appearance can disrupt particular computer vision systems while...
- Connected-component labeling - Connected-component labeling assigns a distinct label to each connected group of foreground pixels in an image.
- Conservative vector field - A conservative vector field is a vector field that can be expressed as the gradient of a scalar potential on its domain.
- Constrained least squares - Constrained least squares finds model parameters that minimize a sum of squared residuals while obeying specified restrictions.
Sources
Michael Isard and Andrew Blake, “CONDENSATION—Conditional Density Propagation for Visual Tracking,” International Journal of Computer Vision, 1998. Primary paper consulted: https://www.cs.cmu.edu/~efros/courses/AP06/Papers/isard-blake-98.pdf
The four-particle calculation and Acorn Vision case are original teaching illustrations, not experimental results from the paper.
Previous: Concurrent logic programming Next: Condition-action rule