Examples
Worked Pop pipelines, from a single ability to multi-stage chains
Each example below is a complete Pop. Pass one to a worker session and it runs against every source you submit. For the attributes these examples use, see Components and Forwarding; for the abilities they name, Models.
from eyepop import EyePopSdk
with EyePopSdk.sync_worker(pop=pop) as endpoint:
result = endpoint.upload("street.jpg").predict()import { EyePop } from '@eyepop.ai/eyepop'
const endpoint = await EyePop.workerEndpoint({ pop }).connect()One ability
The simplest Pop runs a single ability over the whole frame.
from eyepop.worker.worker_types import Pop, InferenceComponent
pop = Pop(components=[
InferenceComponent(ability="eyepop.person:latest"),
])import { PopComponentType } from '@eyepop.ai/eyepop'
const pop = {
components: [
{ type: PopComponentType.INFERENCE, ability: 'eyepop.person:latest' },
],
}Detect, crop, then read
The canonical chain: find vehicles, crop each one, find the plate inside it, crop that, and run OCR on the result. Nothing round-trips to your code between stages.
orientationTargetAngle straightens the cropped plate before OCR, which is what makes angled plates readable. topK keeps only the best plate per vehicle.
from eyepop.worker.worker_types import Pop, InferenceComponent, CropForward
pop = Pop(components=[
InferenceComponent(
ability="eyepop.vehicle:latest",
categoryName="vehicles",
confidenceThreshold=0.8,
forward=CropForward(
includeClasses=["car", "truck"],
targets=[
InferenceComponent(
ability="eyepop.vehicle.license-plate:latest",
topK=1,
forward=CropForward(
orientationTargetAngle=0.0,
boxPadding=0.1,
targets=[
InferenceComponent(
ability="eyepop.text.recognize.landscape:latest",
categoryName="license-plate",
),
],
),
),
],
),
),
])Track people across frames
Tracking hangs off the detector whose objects it should follow. Naming a re-identification ability adds appearance matching, so a person who is briefly occluded keeps their trackId instead of becoming a new object.
Hide a stage the caller never asked for
Facial expression analysis needs a face detector in front of it, but the caller wants expressions attached to people — not a separate list of face boxes. Marking the face detector hidden keeps its output out of the response while still feeding the stage below it, and because it produces object detections, the expression results attach to the enclosing person.
Segment, then trace an outline
A segmentation mask is a per-pixel result. A contour finder turns it into a polygon your application can draw or measure. The finder takes a full forward, because it works on the mask the segmenter produced rather than on a new crop.
Count objects inside one mask
When a segmenter returns a single mask covering many touching objects, a component finder splits it into separate sub-objects. erode shrinks the mask first, breaking the thin bridges between objects that are really separate — here by a kernel 20% of the mask's own size. See How erode and dilate reshape the mask for combining it with dilate.
Find anything, then ask about it
An open-vocabulary ability takes its target through params rather than a fixed label set, and a vision-language ability downstream answers a question about each crop.
Run two analyses side by side
Top-level components run in parallel on the full frame. Use this when a Pop needs two unrelated answers about the same media rather than a chain.
Always return something
crop_with_full_fallback runs the downstream stage on each detection when the parent found objects, and on the whole frame when it found none — so media that defeats the first stage still produces an answer.
Next steps
Components — every component type and attribute
Forwarding — how components chain
Sources and Options — what you can submit to a Pop, and the settings that shape processing
Composable Pops for Python and for Node — constructing these Pops in each language
Last updated