Matchers Overview
The first step in comparing a ground truth and predicted graph is computing a matching between the two graphs. Currently, traccuracy represents a matching between graphs as a matching between the nodes of the graphs - edges are considered matched if the endpoints of the edge are matched. Traccuracy supports four types of matchings, depending on how many nodes in one graph can be matched to one node in the other graph:
one-to-one: every node can be matched to at most one node in the other graphmany-to-one: many ground truth nodes may be matched to one predicted nodeone-to-many: one ground truth node may be matched to many predicted nodesmany-to-many: anything goes! (no implemented metrics support many-to-many at this time)
Many metrics only operate on one-to-one matchings, or support at most one of the “many” mappings. Matchers can operate on the segmentation or directly on the point locations. Below is a table summarizing the implemented matchers, what types of matchings they can produce, and a brief description of behavior and any hyperparameters.
Warning
Unless otherwise noted, matchers are written assuming that you have dense ground truth annotations. The results on sparse annotations may be unpredictable and should be interpreted cautiously.
Matcher |
Matching Type(s) |
Description |
|---|---|---|
|
Given a maximum distance threshold, the matcher will perform hungarian matching on the points in each frame of the ground truth and predicted graphs, minimizing the overall distance while never matching any points with distance greater than the threshold. |
|
|
Given one dataset with segmentation data, this matcher will match nodes from the points-only graph to nodes from the segmentation graph if and only if the point location is within the the segmentation label. |
|
|
Given a minimum overlap threshold, will match the segmentations in each frame of the ground truth and predicted tracks with intersection-over-union greater than or equal to the given threshold. If the one-to-one flag is true, will produce a one-to-one matching by running linear assignment/hungarian matching on the thresholded iou array. |
|
|
A predicted node is matched to a reference node if the computed segmentation covers a majority of the reference segmentation. See https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0144959 for complete details. |