CV Generalized Hough (Shape Template)

Finds an arbitrary SHAPE given as a template image (cv2.createGeneralizedHoughBallard / ...Guil). Where cv2_HoughCircles/HoughLines only fit shapes with a formula, the generalized transform builds an R-table from the template's edge points and lets every scene edge vote for where the shape's reference point must be - so it finds any outline you can draw, and it does it from EDGES, which means partial occlusion and texture/colour changes do not matter (unlike cv2_matchTemplate, which correlates raw pixels). Ballard finds translations only. Guil also searches scale and rotation, which multiplies the accumulator by every scale x angle step - keep the ranges TIGHT or it takes minutes. These are cv2 classes the raw wrappers cannot expose. Zero detections is a valid result (found=false).

By bmad4ever·Created 3 months ago·Updated 2 days ago· 0

image
template

found
positions
scales
angles
votes
bboxes
count

◄methodBallard (position only, fast)►

◄canny_low50►

◄canny_high100►

◄min_distance10►

◄votes_threshold30►

◄dp1.0►

◄levels360►

◄min_scale0.80►

◄max_scale1.20►

◄scale_step0.10►

◄min_angle0►

◄max_angle360►

◄angle_step10.0►

◄position_votes100►

◄scale_votes100►

◄angle_votes10000►

Categoryimage/CV/features

Inputs (18)

Name	Type	Default	Description
image	NPARRAY,IMAGE		Scene to search (converted to grayscale internally; edges are computed inside with the canny_* thresholds). Accepts a ComfyUI IMAGE/MASK directly (frame 0 of a batch) or an NPARRAY. Arithmetic ops (add, multiply, etc.) process the full IMAGE batch when both inputs have the same batch size.
template	NPARRAY,IMAGE		Image of the shape to find. Its edges become the R-table, so a clean, tightly cropped outline works far better than a busy crop. Accepts a ComfyUI IMAGE/MASK directly (frame 0 of a batch) or an NPARRAY. Arithmetic ops (add, multiply, etc.) process the full IMAGE batch when both inputs have the same batch size.
method	COMBO	Ballard (position only, fast)	Ballard: translation only, one accumulator, fast. Guil: also scale and rotation, orders of magnitude slower - narrow min/max scale and angle before using it.
canny_low	INT	501–255	Lower Canny hysteresis threshold used for BOTH the template and the scene edges. Lower it if the shape's outline is faint.
canny_high	INT	1001–255	Upper Canny hysteresis threshold; roughly 2-3x the lower one.
min_distance	FLOAT	101–1000	Minimum distance in pixels between two reported detections - this is what stops one shape being reported as a cluster of near-identical hits.
votes_threshold	INT	301–100000	Ballard only: how many edge points must vote for a position before it counts. LOWER it if nothing is found, raise it if the scene is full of false hits. It scales with the template's edge-pixel count.
dpopt	FLOAT	1.00.5–10	Accumulator resolution divisor: 1.0 = same resolution as the image, 2.0 = half (coarser, faster, more tolerant).
levelsopt	INT	36010–3600	Number of gradient-direction bins in the R-table.
min_scaleopt	FLOAT	0.800.05–10	Guil only: smallest template scale to search.
max_scaleopt	FLOAT	1.200.05–10	Guil only: largest template scale to search.
scale_stepopt	FLOAT	0.100.01–1	Guil only: scale increment. Every extra step costs a full pass.
min_angleopt	FLOAT	00–360	Guil only: smallest rotation in DEGREES to search.
max_angleopt	FLOAT	3600–360	Guil only: largest rotation in degrees.
angle_stepopt	FLOAT	10.00.1–90	Guil only: rotation increment in degrees. 1 degree over the full circle is 360 passes - start coarse.
position_votesopt	INT	1001–100000	Guil only: vote threshold of the POSITION stage.
scale_votesopt	INT	1001–100000	Guil only: vote threshold of the SCALE stage.
angle_votesopt	INT	100001–1000000	Guil only: vote threshold of the ANGLE stage.

Outputs (7)

Name	Type	Description
found	BOOLEAN	True when at least one instance was detected - branch on it with 'Basic data handling: IfElse'.
positions	NPARRAY	Nx2 float32 centres of the detected shapes - feed 'CV Draw Points'.
scales	NPARRAY	(N,) float32 template scale per detection (all 1.0 for Ballard, which does not search scale).
angles	NPARRAY	(N,) float32 rotation in DEGREES per detection (all 0 for Ballard).
votes	NPARRAY	Nx3 int32 accumulator votes (position, scale, angle) - the detection confidence. Sort or threshold on column 0 to keep the strongest hits.
bboxes	BOUNDING_BOX	One {x, y, width, height, score, label} dict per detection: the template's size scaled by 'scales', centred on the position - feed the core 'Draw BBoxes'.
count	INT	How many instances were detected; 0 is valid, not an error.