Composable Pops

EyePop SDK Developer Guide: Using Composable Pops

This guide shows how to build and run Composable Pops using EyePop’s Python SDK. Composable Pops let you chain models and post-processing steps (e.g., object detection → crop → OCR) into reusable pipelines—all defined in Python.

---

🧠 What is a Composable Pop?

A Composable Pop is a flexible visual processing pipeline. Each Pop is made up of components, where each component can:

• Run inference with a specific model

• Crop output and pass it to another component

• Trace objects over time

• Find contours or keypoints

All without needing to train your own model or write any machine learning code.

---

🔧 Installation & Setup

pip install eyepop-sdk requests pillow webui pybars3

Make sure you have:

• An EyePop developer key

---

📦 Composable Pop Examples

Each Pop is defined using the Pop and Component objects.

1. Detect people

Pop(components=[
    InferenceComponent(
        model='eyepop.person:latest',
        categoryName="person"
    )
])

---

2. 2D body pose estimation

Pop(components=[
    InferenceComponent(
        model='eyepop.person:latest',
        categoryName="person",
        forward=CropForward(
            maxItems=128,
            targets=[
                InferenceComponent(
                    model='eyepop.person.2d-body-points:latest',
                    categoryName="2d-body-points",
                    confidenceThreshold=0.25
                )
            ]
        )
    )
])

---

3. Face mesh from people

Pop(components=[
    InferenceComponent(
        model='eyepop.person:latest',
        categoryName="person",
        forward=CropForward(
            maxItems=128,
            targets=[
                InferenceComponent(
                    model='eyepop.person.face.short-range:latest',
                    categoryName="2d-face-points",
                    forward=CropForward(
                        boxPadding=1.5,
                        orientationTargetAngle=-90.0,
                        targets=[
                            InferenceComponent(
                                model='eyepop.person.face-mesh:latest',
                                categoryName="3d-face-mesh"
                            )
                        ]
                    )
                )
            ]
        )
    )
])

---

4. OCR on text regions

Pop(components=[
    InferenceComponent(
        model='eyepop.text:latest',
        categoryName="text",
        confidenceThreshold=0.7,
        forward=CropForward(
            maxItems=128,
            targets=[
                InferenceComponent(
                    model='eyepop.text.recognize.landscape:latest',
                    confidenceThreshold=0.1
                )
            ]
        )
    )
])

---

5. Semantic segmentation with contours

Pop(components=[
    InferenceComponent(
        model='eyepop.sam.small:latest',
        forward=FullForward(
            targets=[
                ContourFinderComponent(
                    contourType=ContourType.POLYGON,
                    areaThreshold=0.005
                )
            ]
        )
    )
])

---

🚀 Running a Pop

from eyepop import EyePopSdk

with EyePopSdk.workerEndpoint() as endpoint:
    endpoint.set_pop(pop_examples["2d-body-points"])  # choose a pop
    job = endpoint.upload("my-image.jpg")             # or use endpoint.load_from(url)

    for result in job.predict():
        print(json.dumps(result, indent=2))

---

🔍 Using Points, Boxes, or Prompts

params = [
    ComponentParams(componentId=1, values={
        "roi": {
            "points": [{"x": 100, "y": 150}],
            "boxes": [{"topLeft": {"x": 10, "y": 20}, "bottomRight": {"x": 200, "y": 300}}]
        }
    })
]
job = endpoint.upload("image.jpg", params=params)

---

🧱 Build Your Own Pops

All components follow this general structure:

InferenceComponent(
    model='your-model:latest',
    categoryName='your-label',
    forward=CropForward(
        boxPadding=0.25,
        targets=[...]
    )
)

Use:

• CropForward for passing regions

• FullForward to process the whole image

• ContourFinderComponent for polygon detection

• ComponentParams to pass prompts or ROI

---

🛠️ Debug Tips

• Log levels can be adjusted using logging.getLogger().setLevel(...)