Skip to content

[Request]: Object detection augmentation with AlbumentationsX and W&B Tables #648

Description

@ternaus

What is the example about?

Would you be open to a self-contained example showing how to apply an AlbumentationsX pipeline to an image, bounding boxes, and class labels together, then inspect the original and augmented annotations in a W&B Table?

A common object-detection data-pipeline failure is that a geometric augmentation changes the image while bounding boxes or their class labels are transformed incorrectly. W&B's interactive bounding-box overlays make this easy to inspect before training.

Proposed scope

The example would:

  • generate a small deterministic detection scene in the notebook, avoiding an external dataset download and data-license dependency;
  • configure AlbumentationsX with bbox_params=A.BboxParams(coord_format="pascal_voc", label_fields=["class_labels"]);
  • apply a small augmentation pipeline to image, bboxes, and class_labels in one call;
  • convert the original and transformed Pascal VOC boxes to W&B's pixel-domain box_data format;
  • log both variants as wandb.Image objects in a wandb.Table;
  • store the AlbumentationsX pipeline configuration in the same W&B run; and
  • omit model training so the example stays focused and quick to execute.

The resulting table would let readers inspect the box overlays and confirm that every transformed box retains the correct class label.

This could be a Colab notebook or a marimo example, whichever format the maintainers prefer.

Dependency details

The example would state that the current public AlbumentationsX package is AGPL-3.0-only. Importing the albumentations module provided by AlbumentationsX also requires an installed PyTorch runtime. PyTorch is intentionally not selected by the AlbumentationsX package metadata because users need the CPU, CUDA, or MPS build appropriate for their environment.

Contribution

I can prepare and run the example end to end, follow the repository's notebook formatting and dependency rules, and add it to the appropriate example index.

Would this scope be welcome? If so, would you prefer Colab or marimo?

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions