TOMTOM17/eps-microplastic-yolo-dataset-05632d7e

EPS and non-EPS mixed-particle dataset

This public dataset contains smartphone-microscopy images of Rhodamine-B-treated particles acquired using 530 nm excitation, a 560 nm long-pass filter, and nominal 10x magnification.

Frozen splits

Split	Images	EPS boxes	Non-EPS boxes	Total boxes
train	23	337	592	929
test	11	218	272	490

The contiguous block IMG_5694 through IMG_5704 is the evaluation split. It was fully held out until the complete pipeline was frozen and was evaluated only once. Do not use images/test or labels/test for fitting, calibration, threshold selection, model selection, or early stopping.

Two development images, IMG_5675 and IMG_5678, had empty label files despite visibly containing particles. They were excluded rather than treated as negative images and are not included here.

Directory structure

images/
  train/       # 23 development/training images
  test/        # 11 evaluation-only images
labels/
  train/       # YOLO labels paired with images/train
  test/        # YOLO labels paired with images/test
protocol/      # immutable split assignment and evaluation ledger
data.yaml

Each label line is:

class_id center_x center_y width height

Coordinates are normalized to image width and height. Class 0 is EPS; class 1 is non-EPS.

Intended use

The boxes support particle-region detection and EPS/non-EPS appearance classification. Because touching particles can be split or merged arbitrarily, exact particle count should be considered diagnostic. The primary scientific endpoint used by the associated pipeline is projected visible EPS area.

This dataset does not provide chemical proof of polymer identity or human-drawn pixel masks. Its images are specific to the documented optical and staining protocol.

Associated model: TOMTOM17/eps-microplastic-frozen-pipeline-05632d7e

Completed source archive SHA-256: 05632d7ee0a7ad2f0c97a1532d6ef097d1d47af41a1c0026f7ea20f732186236