All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog.
- Remove upper limits on dependencies (#38)
- Add module for computing COCO object detection metrics (#36)
- Add Detections class to represent object detections, and load from and save to COCO and parquet formats (#36)
- Drop support for Python 3.8, add support for Python 3.11 (#36)
- Refactor the individual MOT metrics from classes to functions (#36)
- Support for reading/writing to parquet formats (#30)
- Changed internal representation of attributes to use a single numpy array instead of per-frame dict (#30)
- A
py.typed
file to support typing of this library in installations (#25)
- When creating tracks, if
confs
orclasses
are not provided, they are now given a default value, instead of beingNone
. (#24)
- Fix typing for compatibility with Python 3.8 (#22)
- Fix publishing with
hatch
(#21)
-
Fix correct 1-index frame handling for MOT-based files (#20)
-
Fix handling of "outside" detections in CVAT format (#20)
-
Switch to
hatch
for packaging (#20)
- Performance improvement for HOTA metrics using sparse matrices (#19)
- MOT Metrics are now computed using the
MOTMetrics
class, which enables efficient sharing of pre-computed IoU distances across metrics (#17) - Increase minimum Python version to 3.9 (#16)
- Switch packaging system to Poetry (#16)
- More integration tests (though for some reason, there are discrepancies with official results at times) (#12)
- Filtering method for tracks (
filter_by_class
,filter_by_conf
) (#12) - General method for reading CSV file (
from_csv
), reading MOT gt files (from_mot_gt
) (#12)
- Version large files for integration tests with
git-lfs
(#12) - All
Track
class methods for reading CSV files become just a wrapper aroundfrom_csv
(#12)
- HOTA metrics can now be computed, using the original matching algorithm from the paper (gives similar results, but does not correspond to TrackEval's implementation)
Tracks
class, which represents the tracking data (detections), and provides loading functions to read from common MOT data formatscompute_mot_metrics
function, which can calculate MOT metrics (for now CLEARMOT and ID metrics, HOTA coming soon, ALTA in the pipeline)- API documentation hosted on readthedocs
- Pypi package