Classification via instrumented association in multiple-object trackers

The package provides a classifier machinery working on top of an instrumented multi-object tracker fed with identified detections. An example of identified detections are annotations in the simplest case.

The acronym ClavIA stands for classification via instrumented association.

Little explanatory app

Installation

Should be as easy as pip install association-quality-clavia, but if you downloaded the repo, then uv sync standing in the root folder.

Usage

The instrumentation consists in adding an annotation and update IDs to the target objects (tracks) processed in the tracker. The annotation ID is initialized at the target creation time. The update ID is updated after each association procedure.

The classifier procedure (method AssociationQuality.classify) should be called after each tracking step. It is capable of telling apart true positives (TP), false positives (FP), false negatives (FN) and true negatives (TN) if provided with the annotation and update IDs and a supply flag calculated at the current step. The supply flag is easy to get as ann_id in annotation_ids.

The method AssociationQuality.classify returns the object of type BinClass. The class BinClass enumerates TN, FN, FP, TP.

The use of the module will be demonstrated in the package eval-ab-3d-mot. The package eval-ab-3d-mot features the evaluation part extracted from the original AB3DMOT as well as the code to use the association classifier from this package.