1,029 cells: the case for small, honest datasets

The PillWatch student team has captured 159 photographs of a 21-cell pill organizer. Fifty are fully labeled so far, which yields 1,029 labeled cells: 459 full, 570 empty. By modern standards that is not a dataset, it is a rounding error. It has taught the team more than any model has.

Split by photograph, never by cell

The data is split 111/24/24 photographs into train, validation, and test. The unit of the split is the photograph, not the cell. Cells from one photo share lighting, camera position, and pill batch. Split by cell and the test set quietly contains the training set’s conditions, and every reported number inflates. The held-out test photographs are never touched during tuning.

This is the first thing the students had to learn about data: leakage does not look like cheating. It looks like a good result.

The tail is where the truth lives

Most cells are easy. A white pill in a white tray under glare is not.

The team measures camouflage as the color distance between pill and tray, and the labeled set contains only three cells below the perceptual threshold where a pill nearly vanishes into its background. Rare conditions do not show up on their own schedule, so the team runs targeted capture sessions to collect them deliberately instead of waiting.

The hard tail is also where models separate. On 144 deliberately difficult images, a transfer-learned YOLO classifier holds 87.5 percent while the other models drop to roughly 48 to 58 percent and the classic difference-of-Gaussians baseline falls below 40. On the easy 95 percent of the data, those same models look interchangeable.

A dataset you can audit is an argument

The set is condition-stratified on purpose: captures follow a checklist that varies lighting, pill types, and fill patterns, which is why the class balance (459 full, 570 empty) is close to even rather than accidental.

The team is preparing the dataset for public release with a datasheet and a DOI under a CC-BY license. A dataset you can cite is a contribution. A dataset you can audit, with its capture conditions, split rules, and label decisions written down, is an argument that the results mean something.

The data pipeline, model registry, and labels live in the PillWatch data repository. The Raspberry Pi pushes new photographs and labels there automatically; people and training jobs pull.