LibreYOLO1b

LibreYOLO1b

The original YOLOv1 detector ("You Only Look Once", Redmon et al., 2016), repackaged as a LibreYOLO checkpoint for use with the LibreYOLO library.

This is the full 24-convolution YOLOv1 model with its locally-connected and fully-connected head. It is trained on Pascal VOC (20 classes, not COCO) and runs at a fixed 448x448 input (the fully-connected head forbids dynamic shapes).

Source

Derived from the Darknet project (pjreddie/darknet). Darknet is public domain (the "YOLO LICENSE"); the original .cfg architecture and pretrained .weights carry no license obligations.

The pretrained yolov1.weights was published at pjreddie.com/media/files/yolov1.weights (last modified 2016-11-17). That path now returns 404, so the exact file used here was retrieved from the Internet Archive Wayback Machine:

• Archived URL: https://web.archive.org/web/20170124044651id_/http://pjreddie.com/media/files/yolov1.weights
• Size: 777,018,888 bytes
• SHA-256: 624895936c71a41b967fd851a8fbc0fd5c88bcb9f8346b9834ad2cf605826319

The bundled yolov1.cfg reproduces pjreddie's public-domain yolo.cfg with the [connected] output and [detection] num set to the released weights' values (the classic 7x7x30, two-boxes-per-cell head); the LibreYOLO weight reader asserts byte-exact consumption of the .weights file against it.

Modifications

The Darknet .weights binary was converted to a LibreYOLO v1.0 checkpoint (a state-dict mapping into the native LibreYOLO module graph). Learned parameters are unchanged. See weights/convert_darknet_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreYOLO1b.pt")
results = model.predict("image.jpg", save=True)

Metrics

Reference numbers from the paper (VOC2007 test, VOC-style 11-point AP): mAP 63.4. LibreYOLO's validator reports COCO-protocol mAP, which is a different metric; do not compare the two directly.

License

Public domain (Darknet "YOLO LICENSE"). See the LICENSE and NOTICE files in this repository.