LibreECs-pose-obj2coco

NON-COMMERCIAL USE ONLY. Commercial use requires a separate license from INTELLINDUST. These weights are not covered by LibreYOLO's MIT license.

EdgeCrafter S for pose, pretrained on Objects365 and fine-tuned on COCO. This is an explicit opt-in variant; the COCO-only default is LibreECs-pose.pt. This file is not the raw 365-class pretraining model.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreECs-pose-obj2coco.pt")
results = model.predict("image.jpg")

Requires a LibreYOLO version containing EC obj2coco filename support. Existing versions can load a downloaded local checkpoint path if they already support the corresponding EC task.

Source

Official asset: ecpose_s_o3652coco.pth

• Release: edgecrafterv1_o365 (August 14, 2026)
• Source SHA-256: 6d1fb8176199a009afab2d773c15e2cf369e12ac277b3672955ab524b92be62d
• Converted SHA-256: 08f75c7e31323e1ec7e33874ff5e3523d2f6809cafa327e46032c90b8a1e3a03

Copyright (c) 2026 INTELLINDUST INFORMATION TECHNOLOGY (SHENZHEN) CO., LTD.

Modifications and validation

Conversion selects the released EMA when present and adds LibreYOLO metadata. Learned tensors and state-dict keys are unchanged, verified tensor by tensor. Checkpoint schema, factory loading, complete parameter coverage, and a finite 640-pixel CPU forward pass were checked. COCO accuracy and export parity have not been independently re-evaluated for this mirror.

Converter: convert_ec_weights.py.

License

EdgeCrafter License; included verbatim in LICENSE. See NOTICE for attribution and the license history.

The release asset repository still carries Apache-2.0. The main project changed to non-commercial terms on August 20, 2026, after this August 14 release. LibreYOLO's maintainer chose the stricter EdgeCrafter terms for this optional mirror. This is a disclosed redistribution policy, not upstream clarification of the conflicting notices or a claim that earlier Apache grants were revoked.

Objects365's dataset terms state academic-purpose availability, CC BY 4.0 annotations, and separate image restrictions. This repository redistributes model weights, not dataset images.