easydetect lab

English · 한국어

Make an object detector from your own pictures, in a browser: add pictures → draw boxes → train → try it → take it away. On a classroom laptop's CPU with no internet, 100–300 pictures train in 10–15 minutes.

Getting started

A classroom with no internet (Windows): download a zip from the releases, unzip it and double-click start-lab-offline.bat. Nothing is installed and nothing is downloaded.

File	For
easydetect-lab-offline-win64.zip	any laptop (trains on the CPU)
easydetect-lab-offline-win64-intel-gpu.zip	laptops with an Intel GPU (Core Ultra, Arc): training runs on the GPU

For teachers (what to prepare, a lesson plan, troubleshooting), in Korean: docs/classroom.md

A Windows PC with internet: get this repository and double-click start-lab.bat. The first time it installs for a few minutes, then opens the browser. Install Python 3.10–3.13 (3.12 is safest) with Add python.exe to PATH first.

Command line (Windows, macOS, Linux)

git clone https://github.com/themakerrobot/easydetect-lab
cd easydetect-lab
pip install -r requirements.txt
python run.py                  # http://127.0.0.1:8080

Training on a GPU: not needed, but used when there is one. Swap PyTorch for its GPU build once; when the top right shows the GPU's name instead of CPU only, it worked.

rem NVIDIA
.venv\Scripts\python -m pip install torch torchvision --index-url https://download.pytorch.org/whl/cu128
rem Intel GPU (Arc, Core Ultra)
.venv\Scripts\python -m pip install torch torchvision --index-url https://download.pytorch.org/whl/xpu

The EN / 한 button at the top right switches the screens between English and Korean.

How it goes

Follow the five steps along the top of the screen.

1. Add data: pictures, a video (frames are taken at an interval), a zip, or a folder on this PC. A YOLO-format zip from Roboflow and the like comes in with its labels.

2. Draw boxes: drag to draw a box. Class names can be in any language.

   

   Key	Does
   1–9	set the class
   C	copy the previous picture's boxes (handy for a fixed camera)
   Del · Ctrl+Z	delete · undo
   wheel · Space+drag	zoom · pan

   Saving is automatic. Auto-label has a model draw the boxes first, so you only correct them; things COCO knows (an apple, a cup, a person) work from the start. Statistics counts boxes per class and finds missing labels.

3. Train: pick the dataset and a goal, then Start training.

   • Classroom (CPU): picked by itself when there is no GPU; the epochs follow the number of pictures.
   • Fast / Balanced / Accurate: for a PC with a GPU.
   • The rest is set from the data. The curve and the time left update live; afterwards, Evaluate shows the accuracy per class and the pictures with the truth and the prediction drawn over each other.

4. Predict: try it on a picture, a folder, a video or the webcam. The confidence threshold starts at 0.5. For the webcam from another PC, open https://<this PC>:8443.

5. Use the model: download it from the finished run.

   • One ONNX file: one .onnx with the class names and the confidence threshold inside. Anything with onnxruntime runs it — a robot, a Raspberry Pi, a PC (about 0.16 s a picture on a Raspberry Pi 4B). docs/pibo-integration.md
   • OpenVINO IR: fastest on Intel CPUs, GPUs and NPUs.
   • Also best.pt (PyTorch), a folder ready to upload to Hugging Face, and code to copy.

Outlines and keypoints

Pick them under What to label, top left of the labelling screen. Training starts from the same ed-n … ed-x; the dataset's labels add a stage.

• Outlines (segmentation): draw a box and AI (MobileSAM) traces the outline (S), or place the points yourself with P. Training fits the outline model to your data after the box model.
• Keypoints (pose): name the points once (COCO people, 17 points fills them in), then draw a box and click its points in order; Shift+click places a hidden point.

Good to know

• There is no login. By default only this PC can open it. python run.py --host 0.0.0.0 (or start-lab.bat --host 0.0.0.0) opens it to other PCs — and to anyone on that network, so do it only on one you trust.
• The data stays on this PC, in easydetect-lab/ in the folder you start from (data/ inside the offline bundle); --data <folder> moves it.
• One training at a time, in order. Closing the app stops a run, but its best result so far is kept.
• Updating: the app with git pull, the model package with pip install -U "easydetect[train]"; the offline bundle by taking the new zip.
• License: the app, the model code and the weights are all Apache-2.0, free for classes and products alike.

More

• docs/reference.md: every feature in detail, the HTTP API, where files go, security settings, running from source
• docs/classroom.md: preparing a classroom (Korean)
• docs/pibo-integration.md: running on a robot or a Raspberry Pi
• The detection package: easydetect