easydetect hub
Ready-to-run object detectors for 58 public datasets. No PyTorch, no GPU:
pip install easydetect and point it at a file.
- 122 models across 58 datasets, one download each: general objects, everyday objects, people, safety and traffic, animals and environment, agriculture and plants, underwater, aerial and earth, medical and microscopy, documents and screens, industry and electronics, games and video.
- Runs anywhere. ONNX Runtime on any CPU (Windows, macOS, Linux, Raspberry Pi); OpenVINO on an Intel CPU, GPU or NPU.
- Small and fast. ed-s (10M parameters), a real-time DETR: no NMS to tune. 320 px for small devices, with an 8-bit file for a Raspberry Pi; 640 px for accuracy.
- Built on ed-s (Apache-2.0). Each model here is ed-s trained on a public dataset and follows that dataset's terms (in the table and on its page): keep the credit, and check them before commercial use.
- Your own data next. Every model comes with the checkpoint to fine-tune from.
Quick start (no PyTorch)
pip install -U easydetect # 0.7.17 or later
Download a model's folder (its page, or hf download leeyunjai/easydetect-hub --include 'taco/ed-s-640/*' --local-dir hub), then:
from easydetect import Detector
# the file you downloaded: names and threshold inside
model = Detector("taco-ed-s-640.onnx")
# online, a hub name works too (downloads once)
# model = Detector("hub:taco/ed-s-640")
result = model("photo.jpg")[0]
result.show() # or result.save("out.jpg")
print(result.summary()) # [{'name': ..., 'confidence': ..., 'box': {...}}, ...]
Or from the command line, on a picture, a folder, a video or a webcam (source=0):
easydetect predict model=taco-ed-s-640.onnx source=photo.jpg
# easydetect predict model=hub:taco/ed-s-640 source=photo.jpg # online
easydetect hub search=plant # what is here, with scores
pip install easydetect brings numpy, OpenCV, ONNX Runtime and OpenVINO, nothing
more. Only fine-tuning needs PyTorch.
On an Intel NPU or a Raspberry Pi
Detector("taco-ed-s-640.xml", device="NPU") # OpenVINO, .bin beside
Detector("bccd-ed-s-320-int8.onnx", backend="onnxruntime") # 8 bits
# online: Detector("hub:taco/ed-s-640.xml", device="NPU"), Detector("hub:bccd/ed-s-320-int8", ...)
Fine-tune on your own pictures
pip install -U "easydetect[train]" # adds PyTorch
model = Detector("taco-ed-s-640.pt")
# model = Detector("hub:taco/ed-s-640.pt") # online
model.train(data="my_data.yaml", epochs=50, imgsz=640) # YOLO-format folders
Your own dataset: easydetect lab
No model here for what you need? Make it from your own pictures in
easydetect lab, in the browser:
upload pictures, a video or a zip; draw boxes (or let it label them); train (about
15 minutes on a laptop CPU for a class); export to ONNX or OpenVINO. It works
without the internet (offline bundle for Windows). A model trained on your
own pictures is yours to use, products included; a model from here can be its
starting point (upload its .pt).
Without easydetect
Each .onnx runs in plain ONNX Runtime (input: RGB, 0-1, stretched to its size;
outputs: boxes as centre and size, 0-1, and scores per class).
hf download leeyunjai/easydetect-hub --include 'taco/ed-s-640/*' --local-dir hub
What is in each folder
| file | for |
|---|---|
<dataset>-ed-s-<size>.onnx |
ONNX Runtime, any CPU; names and threshold inside |
<dataset>-ed-s-<size>.xml + .bin |
OpenVINO: Intel CPU, GPU, NPU |
<dataset>-ed-s-320-int8.onnx |
8 bits for a Raspberry Pi (320 px only) |
<dataset>-ed-s-<size>.pt |
the checkpoint to fine-tune from (needs PyTorch) |
<model>.mask_decoder.onnx, <model>.pose.onnx |
-seg / -pose: the network beside the detector |
model.json |
classes, size, score, dataset licence |
Models
ed-s. mAP50-95 on each dataset's own val split, COCO's way; in brackets the 8-bit file.
58 datasets: General objects (1) 路 Everyday objects (6) 路 People, safety and traffic (5) 路 Animals and environment (5) 路 Agriculture and plants (7) 路 Underwater (3) 路 Aerial and earth (3) 路 Medical and microscopy (11) 路 Documents and screens (3) 路 Industry and electronics (9) 路 Games and video (5)
| dataset | classes | 320 px (8鈥慴it) | 640 px | licence | |
|---|---|---|---|---|---|
| COCO 2017 General objects 路 detect 路 segment 路 pose |
80 | 0.388 (0.384) | 0.479 | CC BY 4.0 * | |
| chess-pieces-mjzgj (RF100) Everyday objects 路 detect |
13 | 0.793 (0.787) | 0.812 | CC BY 4.0 | |
| coins-1apki (RF100) Everyday objects 路 detect |
4 | 0.803 (0.788) | 0.839 | CC BY 4.0 | |
| currency-v4f8j (RF100) Everyday objects 路 detect |
10 | 0.745 (0.720) | 0.848 | CC BY 4.0 | |
| furniture-ngpea (RF100) Everyday objects 路 detect |
3 | 0.930 (0.927) | 0.938 | CC BY 4.0 | |
| lettuce-pallets (RF100) Everyday objects 路 detect |
5 | 0.699 (0.693) | 0.720 | CC BY 4.0 | |
| marbles (RF100) Everyday objects 路 detect |
2 | 0.501 (0.488) | 0.557 | CC BY 4.0 | |
| Construction-PPE People, safety and traffic 路 detect |
11 | 0.290 (0.292) | 0.304 | CC BY 4.0 | |
| construction-safety-gsnvb (RF100) People, safety and traffic 路 detect |
5 | 0.473 (0.472) | 0.511 | CC BY 4.0 | |
| excavators-czvg9 (RF100) People, safety and traffic 路 detect |
3 | 0.797 (0.793) | 0.794 | CC BY 4.0 | |
| hand-gestures-jps7z (RF100) People, safety and traffic 路 detect |
14 | 0.753 (0.754) | 0.761 | CC BY 4.0 | |
| mask-wearing-608pr (RF100) People, safety and traffic 路 detect |
2 | 0.435 (0.417) | 0.525 | CC BY 4.0 | |
| animals-ij5d2 (RF100) Animals and environment 路 detect |
10 | 0.787 (0.777) | 0.758 | CC BY 4.0 | |
| bees-jt5in (RF100) Animals and environment 路 detect |
1 | 0.353 (0.348) | 0.419 | CC BY 4.0 | |
| cloud-types (RF100) Animals and environment 路 detect |
4 | 0.204 (0.201) | 0.204 | CC BY 4.0 | |
| insects-mytwu (RF100) Animals and environment 路 detect |
10 | 0.647 (0.646) | 0.657 | CC BY 4.0 | |
| TACO Animals and environment 路 detect 路 segment |
60 | 0.144 (0.126) | 0.258 | CC BY 4.0 * | |
| Global Wheat Head 2021 Agriculture and plants 路 detect |
1 | 0.433 (0.428) | 0.486 | CC BY 4.0 | |
| PlantDoc Agriculture and plants 路 detect |
29 | 0.506 (0.501) | 0.564 | CC BY 4.0 | |
| apples-fvpl5 (RF100) Agriculture and plants 路 detect |
2 | 0.537 (0.537) | 0.529 | CC BY 4.0 | |
| cotton-20xz5 (RF100) Agriculture and plants 路 detect |
4 | 0.299 (0.298) | 0.319 | CC BY 4.0 | |
| cotton-plant-disease (RF100) Agriculture and plants 路 detect |
1 | 0.057 (0.056) | 0.066 | CC BY 4.0 | |
| grass-weeds (RF100) Agriculture and plants 路 detect |
1 | 0.378 (0.379) | 0.392 | CC BY 4.0 | |
| leaf-disease-nsdsr (RF100) Agriculture and plants 路 detect |
3 | 0.337 (0.334) | 0.373 | CC BY 4.0 | |
| aquarium-qlnqy (RF100) Underwater 路 detect |
7 | 0.425 (0.417) | 0.502 | CC BY 4.0 | |
| coral-lwptl (RF100) Underwater 路 detect |
14 | 0.107 (0.098) | 0.134 | CC BY 4.0 | |
| fish-market-ggjso (RF100) Underwater 路 detect |
19 | 0.870 (0.868) | 0.872 | CC BY 4.0 | |
| aerial-cows (RF100) Aerial and earth 路 detect |
1 | 0.257 (0.245) | 0.403 | CC BY 4.0 | |
| aerial-pool (RF100) Aerial and earth 路 detect |
5 | 0.307 (0.302) | 0.415 | CC BY 4.0 | |
| aerial-spheres (RF100) Aerial and earth 路 detect |
6 | 0.716 (0.689) | 0.743 | CC BY 4.0 | |
| BCCD Medical and microscopy 路 detect |
3 | 0.630 (0.629) | 0.645 | MIT | |
| abdomen-mri (RF100) Medical and microscopy 路 detect |
1 | 0.552 (0.543) | 0.542 | CC BY 4.0 | |
| acl-x-ray (RF100) Medical and microscopy 路 detect |
1 | 0.744 (0.740) | 0.753 | CC BY 4.0 | |
| axial-mri (RF100) Medical and microscopy 路 detect |
2 | 0.369 (0.355) | 0.422 | CC BY 4.0 | |
| bacteria-ptywi (RF100) Medical and microscopy 路 detect |
1 | 0.045 (0.041) | 0.231 | CC BY 4.0 | |
| bone-fracture-7fylg (RF100) Medical and microscopy 路 detect |
4 | 0.102 (0.095) | 0.143 | CC BY 4.0 | |
| brain-tumor-m2pbp (RF100) Medical and microscopy 路 detect |
3 | 0.463 (0.453) | 0.478 | CC BY 4.0 | |
| cavity-rs0uf (RF100) Medical and microscopy 路 detect |
2 | 0.384 (0.382) | 0.400 | CC BY 4.0 | |
| cells-uyemf (RF100) Medical and microscopy 路 detect |
1 | 0.188 (0.195) | 0.431 | CC BY 4.0 | |
| gynecology-mri (RF100) Medical and microscopy 路 detect |
3 | 0.128 (0.131) | 0.116 | CC BY 4.0 | |
| liver-disease (RF100) Medical and microscopy 路 detect |
4 | 0.391 (0.387) | 0.385 | CC BY 4.0 | |
| activity-diagrams-qdobr (RF100) Documents and screens 路 detect |
19 | 0.354 (0.355) | 0.420 | CC BY 4.0 | |
| digits-t2eg6 (RF100) Documents and screens 路 detect |
10 | 0.727 (0.721) | 0.747 | CC BY 4.0 | |
| document-parts (RF100) Documents and screens 路 detect |
2 | 0.504 (0.495) | 0.522 | CC BY 4.0 | |
| 4-fold-defect (RF100) Industry and electronics 路 detect |
1 | 0.295 (0.268) | 0.390 | CC BY 4.0 | |
| asbestos (RF100) Industry and electronics 路 detect |
4 | 0.360 (0.344) | 0.395 | CC BY 4.0 | |
| cable-damage (RF100) Industry and electronics 路 detect |
2 | 0.402 (0.402) | 0.405 | CC BY 4.0 | |
| cables-nl42k (RF100) Industry and electronics 路 detect |
11 | 0.695 (0.694) | 0.708 | CC BY 4.0 | |
| cell-towers (RF100) Industry and electronics 路 detect |
2 | 0.522 (0.524) | 0.594 | CC BY 4.0 | |
| circuit-elements (RF100) Industry and electronics 路 detect |
31 | 0.015 (0.015) | 0.036 | CC BY 4.0 | |
| circuit-voltages (RF100) Industry and electronics 路 detect |
6 | 0.310 (0.290) | 0.314 | CC BY 4.0 | |
| corrosion-bi3q3 (RF100) Industry and electronics 路 detect |
3 | 0.344 (0.333) | 0.449 | CC BY 4.0 | |
| gauge-u2lwv (RF100) Industry and electronics 路 detect |
2 | 0.364 (0.358) | 0.380 | CC BY 4.0 | |
| apex-videogame (RF100) Games and video 路 detect |
2 | 0.483 (0.483) | 0.577 | CC BY 4.0 | |
| avatar-recognition-nuexe (RF100) Games and video 路 detect |
1 | 0.618 (0.610) | 0.701 | CC BY 4.0 | |
| csgo-videogame (RF100) Games and video 路 detect |
2 | 0.693 (0.690) | 0.705 | CC BY 4.0 | |
| farcry6-videogame (RF100) Games and video 路 detect |
11 | 0.437 (0.429) | 0.453 | CC BY 4.0 | |
| halo-infinite-angel-videogame (RF100) Games and video 路 detect |
4 | 0.611 (0.619) | 0.643 | CC BY 4.0 |
* more terms on the dataset's page (the pictures' own licences).
Segmentation and keypoints
The same detectors with a second network beside them, trained on the
dataset's masks or keypoints, each in a folder of its own
(ed-s-<size>-seg, ed-s-<size>-pose). Download the folder and use the
file, its network beside it, with task=:
Detector("coco-ed-s-640-seg.onnx", task="segment") # .mask_decoder.onnx
Detector("coco-ed-s-640-pose.onnx", task="pose") # .pose.onnx beside
# online, a hub name works too (easydetect 0.7.17 or later fetches both)
# Detector("hub:coco/ed-s-640-seg")
| dataset | task | score | 320 px (8鈥慴it) | 640 px |
|---|---|---|---|---|
| COCO 2017 | segment | mask mAP50-95 | 0.302 (0.298) | 0.339 |
| COCO 2017 | pose | OKS AP | 0.652 (0.650) | 0.663 |
| TACO | segment | mask mAP50-95 | 0.089 (0.092) | 0.137 |
Licence
The models start from ed-s, which is Apache-2.0: easydetect's code and starting weights, from D-FINE, PP-HGNetV2 (PaddleDetection) and, for masks, MobileSAM (NOTICE).
Each model here is ed-s trained on a public dataset, and follows that dataset's terms. Its licence, the credit it asks for and the paper to cite are on its page (click its name): keep the credit with the model, and check the dataset's and its pictures' terms before commercial use.
Made with easydetect.