Keypoints Studio

Keypoints Studio is a desktop tool for building YOLO-style datasets: pose / keypoint labels with Ultralytics models, plus an optional Action Recognition screen for simple bounding-box + class labels (similar in spirit to labelImg).

---

What you can do

Mode	Purpose
Pose (default)	Auto-annotate or manually edit bounding boxes + keypoints, custom keypoint slots, save YOLO pose .txt next to each image.
Action Recognition	Click Switch AR — draw rectangles, pick a class, save YOLO detection .txt (class cx cy w h).

---

Quick start

1. Create conda env and install PyTorch (GPU optional) — see Install.
2. Install from PyPI: python -m pip install cus-keypoints — or from source: python -m pip install -e .
3. Launch: cus-keypoints-gui
4. Pose workflow: pick a root directory → app lists subfolders that contain images → use Next folder / Previous folder → Predict / Preview or Auto-annotate folder → edit → Save label.
5. AR workflow: Switch AR → Open Dir… → select class → Create Box (W) and drag → Save.

---

Screenshots

Add PNGs under images/ in the repo to show them on GitHub:

Preview	File
Main GUI	./images/GUI.png
Keypoint mapping	./images/custom_keypoints.png

If images do not appear, check that the paths and filenames match exactly.

---

Requirements

• Windows 10/11 (other OS may work; scripts below target Windows paths.)
• Python 3.10+ (example uses 3.10)
• Anaconda or Miniconda (recommended)
• NVIDIA GPU + CUDA PyTorch optional — CPU works, slower

---

Install

1. Conda environment

conda create -n torch_gpu python=3.10 -y
conda activate torch_gpu

2. PyTorch with CUDA (optional)

conda activate torch_gpu
conda install -y pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia

Check GPU:

python -c "import torch; print('torch', torch.__version__); print('cuda', torch.cuda.is_available()); print('gpus', torch.cuda.device_count())"

3. Install Keypoints Studio

PyPI (recommended once published):

conda activate torch_gpu
python -m pip install -U pip
python -m pip install cus-keypoints

Editable install from a git clone (folder that contains pyproject.toml):

conda activate torch_gpu
cd /d "C:\path\to\cus_keypoints"
python -m pip install -U pip
python -m pip install -e .

---

Run the GUI

conda activate torch_gpu
cus-keypoints-gui

Console command names are cus-keypoints (CLI) and cus-keypoints-gui (GUI). The Python package you import is still pose_annotator.

---

Pose mode — guide

Project panel

• Switch AR — opens Action Recognition mode (bbox + class only).
• Root directory — choose a folder that contains multiple image subfolders. The app scans immediate child folders that have at least one image, opens the first folder, and shows Folders: i/total.
• Previous folder / Next folder — switch the active annotation folder.
• Recursive — when scanning subfolders for images, include nested images (same flag affects listing inside each folder).

After moving to the next folder

If the current folder already contains keypoints.txt (treated as finished), the whole folder is moved under:

<Root>/Annotation_Done/<folder_name>/

(Annotation_Done is created if needed.)

Preview & tools

Action	How
Zoom	Mouse wheel
Pan	Drag (or middle-drag in some modes)
Next / previous image	Buttons Next / Previous, or keys D / A (global shortcuts in Pose mode)
Run model	Predict / Preview (uses CUDA if available, else CPU)
Save current frame	Save label
Auto-save on image change	Enable Auto save, then use Next / Previous or A / D
Paste bbox from previous image only	Focus preview, Ctrl+V — replaces bbox for selected Person; keypoints unchanged
Paste keypoints from previous image only	Focus preview, Ctrl+B — overwrites keypoints for selected Person
Crosshair + draw new bbox	W — crosshair; drag left button; middle-drag pans
Delete keypoint / bbox	Select item, Delete, or right-click menu

Labeled folder behavior

If a folder contains keypoints.txt:

• Auto-annotate folder is disabled for that folder.
• Predict / Preview does not run the model; it loads image.txt next to each image and shows overlays.

Config files (Pose)

File	Role
pose_annotator/data/custom_classes.txt	Optional extra custom keypoint slots (7. Trachea). Shipped with the package (under site-packages/.../pose_annotator/data/ after pip install). Loaded at startup; merged into the mapping preview (header shown in red).

The repository root data/ folder is only a convenience copy for browsing on GitHub — the running app reads pose_annotator/data/.

Keypoint mapping

Use Mapping… to reorder YOLO keypoints, add/remove custom slots, and rename outputs. The summary under Keypoint mapping reflects the active schema.

---

Action Recognition mode (Switch AR)

Inspired by workflows like labelImg: rectangles + classes, no keypoints.

Item	Detail
Enter / exit	Switch AR (Pose screen) / Switch Pose (AR screen)
Classes	pose_annotator/data/ar_classes.txt — one class per line; class ID = line index starting at 0
Labels	One <image>.txt per image: class_id x_center y_center width height (normalized 0–1)
Also writes	classes.txt in that image folder (class names)
Useful keys	W toggle draw box, A / D prev/next image, Ctrl+S save, Del delete selected box

---

Label formats

YOLO pose (Pose mode)

One row per person:

class_id x_center y_center width height x1 y1 v1 ... xK yK vK

• Coordinates normalized to [0, 1].
• v: typically 0 (missing) or 2 (visible); custom schemas may use more than 17 keypoints.

Docs: Ultralytics pose dataset format

YOLO detection (AR mode)

One row per box:

class_id x_center y_center width height

---

CLI (batch pose auto-annotate)

conda activate torch_gpu
cus-keypoints "C:\path\to\images" --device 0 -v

• --device cpu forces CPU.
• CLI writes labels to a labels directory (see --labels-dir / defaults). The GUI saves *.txt next to each image.

---

Troubleshooting

CUDA missing or Invalid CUDA device

python -c "import torch; print(torch.__version__, torch.cuda.is_available())"

Reinstall CUDA builds from conda (see Install), or use cpu in the GUI device field.

WinError 32 when reinstalling (executable locked)

Close cus-keypoints-gui, then:

python -m pip install -e .

GitHub rejects large .pt files

Do not commit model weights. Keep them local and use .gitignore patterns such as *.pt.

---

Maintainer: publish to PyPI

Distribution name on PyPI: cus-keypoints (the unrelated PyPI project pose-annotator is a different tool).

1. Bump __version__ in pose_annotator/__init__.py.
2. Build and upload:

python -m pip install -U build twine
python -m build
python -m twine upload dist/*

Use a PyPI API token as the password. Prefer trusted publishing from GitHub Actions if you use CI.

---

License

Released under the MIT License — see LICENSE.