Quickstart

This is a pre-release version of Collectra. The API and CLI may change in future releases.

Collectra builds reusable pipelines for extracting data from collection images. Define the tasks and their connections in YAML, run them over your images, and save the resulting crops, text, and classifications with their source relationships.

Installation

Use Python 3.11–3.14. Create a virtual environment and install from PyPI:

python -m venv .venv
source .venv/bin/activate
pip install collectra

On Windows, activate the environment with .venv\Scripts\activate instead.

The default install supports the GUI without Torch. You can inspect existing results, edit text, and view pipelines containing ML tasks without installing their model backends. To run Torch-based inference or training (RF-DETR, DETR, Torchvision/Hugging Face classifiers, or the image orienter), install:

pip install "collectra[ml]"

This extra includes Torch, Torchvision, timm, Transformers, RF-DETR and ensemble-boxes (with its Numba/llvmlite dependencies) for box fusion. It also includes llmloader and its provider integrations for executing LLM tasks. Viewing a pipeline containing LLM tasks does not require llmloader. Model execution still requires compatible hardware and backend wheels.

YOLO tasks and their annotation previews require the optional yolo extra:

pip install "collectra[yolo]"

The default installation excludes Torch, Ultralytics and drawyolo. The YOLO extra installs both; Ultralytics is AGPL-licensed and its terms apply where relevant to combined applications. Making it optional does not remove those obligations. Other backends and the RLSA line detector do not require this extra.

Collectra’s own code remains Apache-2.0. Optional backends and model weights retain their own licences: in particular, check Surya’s version and weights, and the GUI backend selected by pywebview. See the licensing guide before distributing a pipeline or bundled application.

For a source checkout, uv sync and uv run collectra ... use the default installation without Torch. Add --extra ml for the Torch backends or --extra yolo for YOLO. To run the complete test suite, use uv sync --group dev --extra ml --extra yolo followed by uv run --extra ml --extra yolo pytest.

Alternatively, install the development version directly from GitHub:

python -m pip install git+https://github.com/unimelbmdap/collectra.git

Graph rendering also requires the Graphviz system executable (dot) on your PATH. Model-based tasks may download weights on their first run.

Install a pipeline command

If you already have a Collectra pipeline, keep its pipeline.yaml together with any model weights and other files it references. Install a launcher for it:

collectra --pipeline /path/to/pipeline.yaml install my-pipeline
my-pipeline --help
my-pipeline run images --output results
my-pipeline gui results

This creates a command for an existing pipeline; it does not download a pipeline or its model files. The launcher is written to ~/.local/bin by default. Ensure that directory is on your PATH, or choose another destination with --bin-dir. The launcher references the pipeline and Python environment at their current locations, so recreate it if either moves.

You can also run it directly without installing a launcher:

collectra --pipeline /path/to/pipeline.yaml run images --output results

To create your own pipeline, follow the example below.

Design principles and direction

Collectra aims to help researchers and curators turn collection images and documents into structured, searchable data while keeping domain expertise central to the process. Its design is guided by these principles:

  • Compose and reuse tasks. Connect detection, classification, OCR, and LLM tasks into a pipeline suited to your collection. Adapt an existing pipeline or implement a custom task in Python.

  • Keep results connected to their sources. Store artefacts and their relationships in result bundles, with the same format for predictions and training annotations. RO-Crate compatibility is a goal for these bundles.

  • Make human review part of the workflow. Inspect and correct predictions before using them as research data or training examples.

  • Improve models with reviewed examples. Fine-tune supported models on corrected annotations and repeat the predict, review, and train cycle.

  • Make annotation effort count. Planned clustering and embedding-based browsing will help reveal similar items and prioritise diverse examples.

Many of these features are still to come. Today, pipelines are defined in YAML, the GUI supports reviewing and editing results, and supported models are trained through the CLI. The visual pipeline builder, training from the GUI, RO-Crate-compatible packaging, interactive clustering, and diversity-based annotation ordering remain planned work. Landmark/pose detection, segmentation, and oriented bounding boxes are also part of the intended task range.

See the design principles and development direction for more detail.

Build your first pipeline

A pipeline connects tasks (operations) through named artefacts (data). For example, a line detector takes an Image and produces ImageCrop artefacts. An OCR task can consume those crops and produce Text artefacts. Tasks declare these connections using input and output names.

Create a working directory with the following layout:

my-project/
    pipeline.yaml
    images/
        page001.png
        page002.png

Save this as pipeline.yaml:

collectra_pipeline_metadata:
  name: printed-pages
  ext: collectra
  version: "1.0"

detect_lines:
  type: collectra.LineDetectorRLSA
  input: page
  output: lines
  deskew: true

This pipeline detects lines of printed text without loading a pretrained model. Run the commands below from my-project.

The metadata defines the pipeline’s name, version, and result-directory suffix: ext: collectra produces directories such as page001.collectra. The page node receives each input image. detect_lines is the task name used in CLI commands; lines is the output artefact name. Collectra infers the Image input and ImageCrop output types from the task, so separate page and lines definitions are not needed.

Every task needs a type identifying an importable Python class. Additional keys, such as deskew or model, configure that task. Input and output names can be strings or lists. Connecting another task to lines makes it run after the detector. Keep these connections acyclic.

Run a pipeline

Process one image or a directory of images:

collectra --pipeline pipeline.yaml run images/page001.png --output results
collectra --pipeline pipeline.yaml run images --output results

You can also supply multiple input paths. --pipeline accepts either a YAML file or a directory containing pipeline.yaml. Without --output, result directories are created alongside the input images.

Discover commands and task-specific options with --help:

collectra --pipeline pipeline.yaml --help
collectra --pipeline pipeline.yaml run --help
collectra --pipeline pipeline.yaml task --help
collectra --pipeline pipeline.yaml task detect_lines run --help

Add --verbose to see detailed processing logs, or --render to generate workflow diagrams while processing. The commands available under task depend on the tasks defined in your pipeline.

Inspect and reuse results

Each input gets a result directory:

results/
    page001.collectra/
        page001.png
        results.yaml
    page002.collectra/
        page002.png
        results.yaml

results.yaml records the pipeline metadata, artefact types, IDs, data, and parent relationships. Crops record their bounds and source image; they need not be separate image files. Other tasks may create additional files in the same result directory.

Open results in the desktop GUI to inspect the image, crops, and other outputs:

collectra --pipeline pipeline.yaml gui results/page001.collectra
collectra --pipeline pipeline.yaml gui results

The GUI requires a desktop environment and a working pywebview backend. It opens existing results; run the pipeline first to create them.

Run just one task on existing results:

collectra --pipeline pipeline.yaml task detect_lines run results/page001.collectra

To start at a task and continue through its downstream tasks, use:

collectra --pipeline pipeline.yaml run results/page001.collectra --task detect_lines

The task’s inputs must already exist in the results. A task-level run runs only that task; a pipeline-level run --task ... also runs its downstream tasks. Running on an existing result directory updates its outputs.

Add OCR and join the text

To extend the example, append these tasks to pipeline.yaml:

read_lines:
  type: collectra.OCRSuraya
  input: lines
  output: line_text

join_text:
  type: collectra.ConcatenateText
  input: line_text
  output: transcription
  separator: "\n"

The data flow is now page → detect_lines → lines → read_lines → line_text → join_text → transcription. OCR runs on each crop; ConcatenateText collects the resulting text artefacts and joins them into one transcription.

OCRSuraya is the Collectra class name for the Surya OCR backend. It requires the surya-ocr package in your environment and downloads its model weights on first use. The initial line-detection example does not require this backend.

After installing the OCR backend, process new images with the same run command, or start OCR on previously detected lines:

collectra --pipeline pipeline.yaml run results/page001.collectra --task read_lines

Choose a task

Built-in task types include:

  • collectra.LineDetectorRLSA: detect printed-text lines using image processing.

  • collectra.LineDetectorSurya: detect text lines with Surya models.

  • collectra.OCRSuraya: recognize text in images or crops.

  • collectra.ConcatenateText: combine text artefacts.

  • collectra.ObjectDetectionYOLO, collectra.ObjectDetectionDETR, and collectra.ObjectDetectionRFDETR: produce labelled object crops.

  • collectra.ImageClassifierYOLO, collectra.ImageClassifierTorchvision, and collectra.ImageClassifierHuggingFace: classify images or crops, producing Link artefacts that retain a reference to the classified image.

  • collectra.LLM: process inputs with a configured language model.

Choose model weights trained for the objects or classes in your collection. Task-specific backends, model configuration, and credentials may be needed; use the documentation and the configured task’s --help to inspect its options.

Train a classifier

Trainable tasks expose their own train command. For example, save this separate pipeline as classifier/pipeline.yaml:

collectra_pipeline_metadata:
  name: specimen-classifier
  ext: collectra
  version: "1.0"

classify:
  type: collectra.ImageClassifierTorchvision
  model: resnet18
  input: image
  output: [Pollen, Spore]

Prepare annotated *.collectra result directories under training-data. For classification, each label is a Link named after an output class and linked to its source image or crop. Set the result metadata’s partition to validation for held-out examples. A minimal annotated results.yaml is:

collectra_results_metadata:
  partition: validation

image:
  type: collectra.Image
  id: image-001
  data: specimen.png

Pollen:
  type: collectra.Link
  id: label-001
  parents: image-001

Place specimen.png alongside this file. Provide training examples for each class as well, using a different partition such as training. Detection tasks instead train from labelled ImageCrop annotations.

collectra --pipeline classifier/pipeline.yaml task classify train --help
collectra --pipeline classifier/pipeline.yaml task classify train training-data \
    --validation validation --epochs 10 --batch 16 --workers 0

Training saves the best checkpoint into the pipeline directory and updates the task’s model entry in pipeline.yaml. Training logs are retained by default; --no-keep-log removes the training directory after saving the model. For this Torchvision classifier, Weights & Biases logging is off by default. Add --wandb to opt in.

Use the trained pipeline on new images:

collectra --pipeline classifier/pipeline.yaml run images --output classified

Use a pipeline from Python

The same YAML configuration can be loaded and run from Python:

from collectra import Collectra

pipeline = Collectra.from_file("pipeline.yaml")
pipeline.cli_run(["images"], output="results")

# Start at a particular task using existing artefacts.
pipeline.cli_run(
    ["results/page001.collectra"],
    task="detect_lines",
    single=True,
)

# Requires Graphviz's dot executable.
pipeline.render("pipeline-graph", render=True)

cli_run resolves file and directory inputs in the same way as the CLI. For an explicit list of files, you can also use pipeline.run(files=["images/page001.png"], output="results").

For more task-specific guidance, see the Collectra documentation.