Tracking Parameters¶
The following parameters control the behavior of the tracking
algorithms used in slurpy.
Snake (Active Contour) Parameters¶
In computer vision, a “snake” is a flexible spline that automatically deforms to map onto edges. It balances internal forces (stiffness/elasticity) against external forces (image features).
alpha: 0.1 (Continuity / Elasticity)¶
This parameter controls the stretchiness of the snake, acting like a rubber band.
Function: It penalizes the contour if anchor points get too far apart or cluster too closely. It maintains even spacing.
High values: The snake acts like a tight rubber band. It resists stretching and may pull away from sharp corners to keep points close.
Low values: The snake is loose. Points can stretch far to grab distant edges, but may bunch up in areas with high gradients.
lambda1: 0.5 (Smoothness / Energy Weighting)¶
This parameter controls the stiffness of the snake, acting like a flexible wire.
Function: It resists bending. In this implementation,
lambda1balances the energy components to dictate overall curve smoothness.High values: The snake acts like a stiff wire. It resists sharp turns, resulting in smooth, sweeping curves that may ignore jagged details.
Low values: The snake acts like a flexible string. It easily bends into sharp corners to perfectly hug every tiny variation in a boundary.
band_penalty: 10.0 (Boundary Constraint)¶
This is a spatial constraint used to prevent the contour from wandering away from the target.
Function: It acts as a penalty (a “wall”) if points venture outside a defined target band around your initial anchor points.
High values: Points trying to track edges outside the safe zone are heavily penalized, forcing the snake to stay near your initial guess.
Low values: The snake is allowed to roam freely across the image to find the strongest edges, even if they are far from the starting position.
Particle Filter Parameters¶
The particle filter uses statistical motion modeling to generate and evaluate perturbed contours (particles) over sequential frames.
num_particles: 50 (Particle Count)¶
This parameter defines the number of hypothesized contours.
Function: It specifies how many unique particle variants to generate during evaluation.
High values: More particles increase the chances of finding the optimal boundary but slow down processing time.
Low values: Faster processing, but higher risk of losing the tracked object during rapid movement.
percent_var: 0.98 (Shape Model Variance)¶
This parameter handles the proportion of statistical variance kept in the Principal Component Analysis (PCA).
Function: Controls how much characteristic shape variation is permitted from the learned active shape model.
High values: Retains more subtle shape deformations, allowing flexibility for complex bounds.
Low values: Forces the tracker to adhere strictly to the most dominant mean shapes, rejecting unusual contour forms.
noise_scale: 1.0 (Tracking Spread)¶
This parameter controls the random noise multiplier.
Function: Simulates potential state variations by applying Gaussian noise matched to the contour shape.
High values: Particles spread out further, enabling recovery from sudden jumps or large frame-to-frame movements.
Low values: Particles stay tightly clustered around the base contour, providing stable tracking for slow motions.
Quick Tuning Guide¶
Problem |
Solution |
|---|---|
Tracking is too jagged/noisy |
Increase |
Points cluster or leave gaps |
Increase |
Snake snaps to wrong object |
Increase |
Particle tracker loses object |
Increase |
Tracker runs too slowly |
Decrease |