KvasirCapsule Loader

An inofficial abstraction for the KvasirCapsule dataset that aims to provide the following features:

• Convenient abstraction for KvasirCapsule as a Pytorch Dataset with augmentations
• Patient-separating dataset splits that are truely separate and reproducible
• Auto-download KvasirCapsule dataset
• Canonical representation of finding categories, finding classes, folder structures and bounding boxes
• Unit tests, type hints for type checking and error handling

Rationale

Convenience: Loading a dataset for an experiment should be as simple as issuing a single function call.

Proper Data Splits: The KvasirCapsule dataset itself features an official two-fold split. However, this split seems to have patient overlap, which may lead to overfitting when used in a training setup. Further, a two-fold split is not enough for a training/validation/test setup, k-fold cross-validation, or out-of-distribution detection.

Modern Python tooling: Enables a smooth workflow and up-to-date dependencies.

Error and type checking: Canonical representations of classes, categories and bounding boxes as dedicated types allow for proper error handling and avoid common pitfalls in data processing pipelines.

Getting Started

The latest release of this project can be installed through pip:

pip install kvasircapsuleloader

Once installed, including KvasirCapsule as a pytorch dataset is as simple as running:

...
from kvasircapsuleloader import KvasirCapsuleDataset

dataset = KvasirCapsuleDataset()
dataloader_train = DataLoader(dataset.train())
dataloader_val = DataLoader(dataset.val())
dataloader_test = DataLoader(dataset.test())

Please note that this call will automatically download the KvasirCapsule dataset from the OSF repo if it is not available yet.

Roadmap

TODOs:

• Splits for k-fold cross-validation
• Splits for OOD-detection (held-out training set)
• Visualization utilities
• Allow user to select only samples with bounding boxes
• Selection criteria, only include selected classes
• Unit tests for all relevant modules
• Download unlabelled videos

Acknowledgement

This project provides an abstraction for the KvasirCapsule Dataset by Smedsrud et al., which is licensed under CC-BY-SA 4.0. Please refer to the Kvasir Capsule GitHub Repository for terms of use and the corresponding BibTeX entry.