Modelpublic

LibreYOLO/LibreOVDEIMm

sourceHugging Facecc-by-nc-4.0updated 2mo agoView on Hugging Face

0likes60downloads

Model Card

LibreOVDEIMm

OV-DEIM-M real-time open-vocabulary detector (20M params, ViT-tinyplus (DEIMv2 distilled) backbone) repackaged for LibreYOLO's LibreOpenVocab tier, bundled with the MobileCLIP-B(LT) text tower for online prompt encoding.

Source

Detector converted from wleilei/OV-DEIM at commit dfbf394672407b7f837ec08e7d68e8127548b254 (checkpoint ovdeim_m.pth, trained on Objects365v1 + GoldG). Text encoder taken from apple/MobileCLIP-B-LT-OpenCLIP.

Paper: OV-DEIM: Real-time DETR-Style Open-Vocabulary Object Detection with GridSynthetic Augmentation (arXiv 2603.07022). Upstream reports 109 FPS on a T4 with TensorRT at 640 input.

Modifications

State-dict key remapping and merge only (module. prefix stripped, the training-only denoising embedding dropped, the MobileCLIP text tower stored under text_encoder.*). Learned parameters are unchanged. See weights/convert_ovdeim_weights.py in the LibreYOLO source repository.

LibreYOLO's port reproduces the upstream paper's zero-shot COCO val2017 result through this artifact (45.9 AP for the L size).

Usage

python

from libreyolo import LibreOpenVocab

model = LibreOpenVocab("ov-deim-m")
model.set_classes(["person", "a red backpack", "forklift"])
results = model.predict("image.jpg")

License

Layered, most restrictive governs the combined artifact:

—Detector weights: CC BY-NC 4.0 (non-commercial, attribution; redistribution and format conversion permitted by upstream's MODEL_LICENSE).
—OV-DEIM code lineage: Apache-2.0.
—MobileCLIP-B(LT) text tower: Apple Machine Learning Research Model license, research use only.
—DINOv3-derived backbone: DINOv3 License.

See `NOTICE` for attribution details.