andreribeiro87/kvasir-seg-augmented

Kvasir-SEG Augmented

A preprocessed and augmented version of the Kvasir-SEG polyp segmentation dataset, used to train the models in the UNet-A benchmark.

Dataset Description

Kvasir-SEG is a publicly available dataset of 1,000 colonoscopy images annotated with binary polyp masks, introduced by Jha et al. (2020). This version adds offline data augmentation to the training split and provides fixed validation and test partitions for reproducible benchmarking.

Splits & Sizes

Split	Images	Notes
train	4,800	Original 800 images × 6 (1 original + 5 augmented copies), 256 × 256
validation	100	Original validation split from Kvasir-SEG, 256 × 256
test	100	Original test split from Kvasir-SEG, 256 × 256

Augmentation Strategy (training split only)

The original Kvasir-SEG provides pre-defined train/validation/test splits (800/100/100). Each original training image produces one resized copy (256 × 256) plus five augmented variants:

• Random horizontal flip (p = 0.5)
• Random vertical flip (p = 0.5)
• Random rotation ±30° (p = 0.5)
• Random brightness / contrast / saturation jitter ±20 % (p = 0.5)

Augmentation was performed with a fixed seed (42) for full reproducibility.

Features

{
    "image": Image(RGB),   # colonoscopy frame
    "mask":  Image(L),     # binary polyp mask (0 = background, 255 = polyp)
}

Usage

from datasets import load_dataset

ds = load_dataset("andreribeiro87/kvasir-seg-augmented")

sample = ds["train"][0]
image = sample["image"]   # PIL.Image RGB
mask  = sample["mask"]    # PIL.Image L (grayscale)

Source

• Original dataset: Jha, D. et al. "Kvasir-SEG: A Segmented Polyp Dataset." MMM 2020. [paper] · [original HF]
• This version: created as part of the UNet architecture benchmark — see andreribeiro87/unet3plus-efficientnet-kvasir-seg for the best-performing model.

License

The underlying Kvasir-SEG images are distributed under CC BY 4.0.