LibreReID-osnet

OSNet-AIN person re-identification weights for LibreYOLO's Deep OC-SORT tracker (model.track(source, tracker="deepocsort")). Downloaded automatically on first use to ~/.cache/libreyolo/reid/.

Files

File	Size	SHA-256
osnet_ain_x0_25.pt	1.0 MB	ce171fe160b3608f5e4c19489774991419be965b1d6f4bdccc4b4cfd2ef95347
osnet_ain_x0_5.pt	2.7 MB	510bcebae21bd0c0fcc7df388e97d2f687a9ee4befa4394d6fb1fb19aac0bce2
osnet_ain_x0_75.pt	5.4 MB	57b31d7f806edac586540e08e98c589f0010ad7876dacf4af013deaa284dd26d
osnet_ain_x1_0.pt	8.9 MB	34c24e98b6b70c8b62480f846fd0d581aa2fd1535bc0276aecf1f10430b731d1

osnet_ain_x0_25 is the LibreYOLO default. All files are plain PyTorch state dicts producing L2-normalized 512-d embeddings.

Provenance and license

• Network: OSNet-AIN, ported to LibreYOLO from Torchreid (MIT). The LibreYOLO port is state-dict compatible and bit-exact against upstream (tests/unit/test_reid.py).
• Weights: converted unchanged from the Torchreid model zoo multi-source (MS+D+C) OSNet-AIN checkpoints, released under the repository's MIT license. Conversion script: weights/convert_osnet_reid_weights.py in the LibreYOLO repository (strips the classifier head, keeps feature layers, verifies strict load).
• Training data: the upstream checkpoints were trained by the Torchreid authors on person re-identification research datasets (MSMT17, DukeMTMC-reID, CUHK03). Those datasets carry research-oriented terms; the weights themselves are distributed under MIT by the upstream author. Review your own use case if you deploy person re-identification in production.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreYOLO9t.pt")
for result in model.track("video.mp4", tracker="deepocsort"):
    print(result.track_id)