Datasets and metadata¶
A dataset is the unit everything else hangs off. Images, labels, annotations, quantification profiles, calibration defaults and membership all belong to exactly one dataset.
This is deliberate. Label names are unique dataset-wide, so a label means one thing everywhere inside a dataset, and a measurement can be compared across every image in it without asking whether two people meant the same thing by "polyp".
Images¶
Images are uploaded into a dataset and browsed in its gallery. Each image carries:
- the pixel data itself
- typed per-image metadata you define
- a calibration, if one has been set — see Calibration
- the annotations drawn on it — see Objects and annotations
Typed metadata¶
Metadata fields are typed rather than free-form strings, which is what makes them useful for filtering and grouping: a depth is a number you can range-filter, a survey site is a category you can group by, a capture date is a date you can sort on.
In practice this is what lets you ask questions of the analysis rather than of individual images — "mean colony area by site" only works if "site" is a field and not a naming convention buried in filenames.
dataset-metadata-filterGallery filtered and grouped by a typed metadata field.
6–8 s · dark theme · 1280×800
Practical advice¶
- One dataset per measurement question, not per folder on disk. If two sets of images will never be compared to each other, they are two datasets.
- Decide the metadata fields before bulk upload. Adding a field later means backfilling it.