There are several reference annotations available. The two main categories, available for the training datasets but kept in secret for the challenge datasets, are:
- Reference segmentation annotations consisting of reference cell masks used to evaluate the segmentation performance of competing methods;
- Reference tracking annotations consisting of cell markers that are interlinked between frames to form cell lineage trees used to evaluate the detection and tracking performance of competing methods.
Furthermore, we distinguish the corpora containing reference annotations based on their origin:
- Absolute truth corpus (ground truth) containing exact reference annotations, available only for simulated datasets;
- Gold-standard corpus (gold truth) containing human-made reference annotations, obtained as a consensual or majority opinion of several human experts;
- Silver-standard corpus (silver truth) containing computer-generated reference annotations, obtained as the majority opinion over the results of several automatic analysis methods submitted by former challenge participants.
For real datasets, the following reference annotations exist:
- Gold reference tracking annotation (gold tracking truth) with complete cell instance coverage (unless explicitly stated) but poor cell region information;
- Gold reference segmentation annotation (gold segmentation truth) with very limited cell instance coverage but good cell region information;
- Silver reference segmentation annotation (silver segmentation truth) with good cell instance coverage but limited cell region information; so far available for nine datasets.
Figure 1: An example of a real dataset and its annotations: original image data (top left), gold tracking truth (top right), gold segmentation truth (bottom left), silver segmentation truth (bottom right).
For simulated datasets, only ground truth reference annotation exists. It consists of the exact segmentation masks for all cells, which are simultaneously used as tracking markers too and accompanied by the exact cell lineage tree information.
Figure 2: An example of a simulated dataset and its annotation: original image data (left), ground truth (right).
More Details and Download
The detailed annotation guidelines are described in the following document:
Both the original image data and all available reference annotations can be downloaded as a single ZIP archive for each training dataset. The challenge datasets are distributed as the ZIP archives too, but without reference annotations. Individual files in the ZIP archives were named and created following the conventions described in:
Naming and file content conventions.pdf
For download, please continue to 2D+Time Datasets or 3D+Time Datasets.