easydetect weights — D-FINE for pip install easydetect

Real-time object detectors, ready to run: ONNX for ONNX Runtime on any CPU, OpenVINO IR for an Intel CPU, GPU or NPU, and the PyTorch checkpoint to fine-tune from. These are the files the easydetect package downloads on first use — you never have to fetch them by hand.

Apache-2.0 end to end: the code, the COCO weights, and the models fine-tuned here.

pip install easydetect          # OpenVINO and ONNX Runtime, no PyTorch

from easydetect import Detector

model = Detector("dfine-s")                 # downloads dfine-s/ from this repo once
results = model("photo.jpg", conf=0.5)
results[0].save("result.jpg")

for r in model.predict(0, stream=True, show=True):   # webcam; q or Esc quits
    pass

No NMS step to tune: D-FINE predicts its set of boxes end to end; the package drops the rare second box on one object (IoU > 0.7) and handles resize, decode and drawing.

COCO models — 80 classes

folder	backbone	decoder	params	COCO mAP50-95
dfine-n	HGNetv2-B0	3 layers	4M	42.8
dfine-s	HGNetv2-B0	3 layers	10M	48.5
dfine-m	HGNetv2-B2	4 layers	19M	52.3
dfine-l	HGNetv2-B4	6 layers	31M	54.0
dfine-x	HGNetv2-B5	6 layers	62M	55.8

Every folder holds the same files:

file	what it is
<name>.xml + <name>.bin	OpenVINO IR — what Detector("<name>") runs on OpenVINO
<name>.onnx	the same network for ONNX Runtime — the light install, e.g. a Raspberry Pi
<name>.pt	PyTorch checkpoint — the starting point for model.train(...)
labels.txt	class names, one per line

dfine-s at 640 × 640 on a Core Ultra 5 250K Plus, whole pipeline:

device	latency
CPU	36 ms (28 FPS)
NPU	38 ms (26 FPS) — the CPU stays free

More in performance.

Fine-tuned models — models/

Detectors trained on one job, each with its own README: classes, scores, the data it learned from, and how it was trained.

folder	finds	base	val mAP50-95

Use one with huggingface_hub:

from huggingface_hub import snapshot_download
from easydetect import Detector

root = snapshot_download("leeyunjai/easydetect", allow_patterns="models/<name>/*")
model = Detector(f"{root}/models/<name>/best.xml", device="AUTO")   # CPU · GPU · NPU
model.predict("photo.jpg", save=True)

Train your own

pip install "easydetect[train]"

from easydetect import Detector

model = Detector("dfine-s")                        # starts from the COCO weights above
model.train(data="data.yaml", epochs=50, imgsz=640)
model.export(format="openvino")                    # best.xml + best.bin + labels.txt

data.yaml is the common images/ + labels/ layout — a Roboflow export works as it is. Or do it all in a browser: the easydetect lab collects and labels images, trains, shows the numbers, and writes the upload folder for models/ with its README filled in from the run.

Layout

dfine-n/   dfine-n.xml  dfine-n.bin  dfine-n.pt  labels.txt
dfine-s/   …
dfine-m/   …
dfine-l/   …
dfine-x/   …
mobile_sam/  encoder.onnx  decoder.onnx  LICENSE    the segmenter (task="segment")
pose/        pose-s.onnx                           the keypoint model (task="pose")
models/
  <name>/   best.xml  best.bin  labels.txt  README.md

mobile_sam/ is MobileSAM (Apache-2.0, ChaoningZhang/MobileSAM) exported to ONNX by tools/convert_sam.py: given the boxes a detector found, it outlines what is inside each.

pose/ is easydetect's own keypoint network (Apache-2.0, 3.7 M parameters), trained by tools/train_pose.py on COCO 2017 person keypoints (CC BY 4.0) from the dfine-s backbone: given a person box, it places COCO's 17 body keypoints. OKS AP on COCO val2017, reading each person mirrored too (the package's default): 0.637 with the labelled boxes, 0.588 behind dfine-s.

Keep these names: the package builds its download URLs from them (<repo>/resolve/main/<name>/<name>.xml). To host a copy elsewhere, mirror the same layout and point $EASYDETECT_ASSETS_URL at it.

License and credit

• The COCO weights are the official D-FINE checkpoints by Yansong Peng et al., released under Apache-2.0 at Peterande/D-FINE, converted unchanged. Only the COCO-trained checkpoints are used. COCO annotations are CC BY 4.0.
• The package and the fine-tuned models are Apache-2.0. Each model's README says where its training images came from; those images keep their own license.

@misc{peng2024dfine,
  title         = {D-FINE: Redefine Regression Task in DETRs as Fine-grained Distribution Refinement},
  author        = {Yansong Peng and Hebei Li and Peixi Wu and Yueyi Zhang and Xiaoyan Sun and Feng Wu},
  year          = {2024},
  eprint        = {2410.13842},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}