Skip to content

Pre-Annotation Pipeline

This guide explains how to create, customize, test, and deploy a pre-annotation pipeline using pxl-pipeline cli with the pre_annotation template.

These pipelines apply an existing model (e.g. YOLOv8, GroundingDINO) to automatically annotate a dataset.

1. Initialize your pipeline

pxl-pipeline init my_preannotation_pipeline --type processing --template pre_annotation

This generates a pipeline folder with standard files. See project structure for details.

2. Customize your pipeline

steps.py

Contains the process() step where the model is applied to the dataset.

@step
def process(picsellia_model: Model, picsellia_dataset: CocoDataset):
    output_coco = process_images(
        picsellia_model=picsellia_model,
        picsellia_dataset=picsellia_dataset,
        parameters=parameters,
    )
    ...
    return picsellia_dataset

utils/processing.py

Implements process_images() where you:

  • Apply the model to each image
  • Generate bounding boxes or segmentation masks
  • Format the output in COCO

You can modify this logic to use a different model (e.g., GroundingDINO) or post-process detections.

utils/parameters.py

Define your custom parameters (e.g. threshold):

self.threshold = self.extract_parameter(["threshold"], expected_type=float, default=0.1)

Learn more in Working with pipeline parameters.

3. Manage dependencies with uv

This template uses uv for dependency management. Dependencies are declared in pyproject.toml and resolved automatically.

To add packages:

uv add opencv-python --project my_preannotation_pipeline
uv add git+https://github.com/picselliahq/picsellia-cv-engine.git --project my_preannotation_pipeline
See dependency management with uv for full details.

4. Test locally

pxl-pipeline test my_preannotation_pipeline

You’ll be prompted for:

  • input_dataset_version_id
  • model_version_id

A new folder will be created under runs/, storing config and results.

To reuse the same folder and avoid re-downloading assets, use:

pxl-pipeline test my_preannotation_pipeline --reuse-dir

See how runs/ work for more details.

5. Deploy to Picsellia

pxl-pipeline deploy my_preannotation_pipeline

This will:

  • Build and push the Docker image

  • Register the pipeline in Picsellia under Processings → Dataset → Private

See deployment lifecycle.

Make sure you’re logged in to Docker before deploying.