Tutorials

Launch F-Annote

Run F-Annote.exe from the folder you downloaded. The animated splash screen loads your hardware info and opens the main window.

Click "New Project"

Hit the "New Project" button in the left sidebar. The 3-step project wizard opens.

Step 1 - Name your project

Enter a project name. Only letters, numbers, underscores, and hyphens are allowed - no spaces. Example: "cs2_detection" or "car-detector".

Step 2 - Define your classes

Add the object classes you want to detect, e.g. "Car", "Person", "Truck". You can add more classes later from inside the annotator. You need at least one class to continue.

Step 3 - Training settings

Choose a base YOLO model (yolov8n is fastest, yolov8x is most accurate), image size (640 is the standard), and your compute device. If you have an NVIDIA GPU it will appear in the list - select it.

Your project is ready

Click Finish. F-Annote creates the project folder structure and opens it automatically. Add your images from the overview panel on the main screen, or click "Open Annotator" to start.

Open the Annotator

With a project open, click "Open Annotator" in the tools panel. The annotator window opens with your images loaded in the thumbnail strip at the bottom.

Add more images by drag & drop

Need to add images while you work? Just drag image files (or a whole folder) straight onto the annotator window and drop them. They are copied into your dataset, deduped by name, and appear in the thumbnail strip instantly - no need to go back to the main screen.

Draw a bounding box

Left-click and drag on the canvas to draw a box around an object. Release the mouse to confirm. The box is saved automatically if autosave is on.

Select the right class

The active class is shown in the class panel on the left. Press number keys 0–9 to instantly switch between classes. If you have a box selected, pressing a number key also changes that box's class immediately.

Move and resize boxes

Click a box to select it (it turns highlighted). Drag the box body to move it. Drag any of the 8 corner/edge handles to resize. Minimum box size is 8 pixels.

Delete a box

Select a box and press Delete or Backspace to remove it. Right-click a box to get a context menu with Delete and Change Class options.

Navigate between images

Click any thumbnail at the bottom, or use the Next / Previous buttons in the toolbar. You can also press the left and right arrow keys to step through images. Labels save automatically before moving.

Filter unannotated images

Click the "Unannotated" button in the toolbar. It turns amber when active and shows how many images still need labels, e.g. "5 left". The filter clears automatically when you annotate the last remaining image.

Two modes - your model or a COCO pretrained model

You do not need a trained model to start. F-Annote includes COCO pretrained base models (yolov8n, yolov8s, yolo11n, yolo11s) that can detect 80 common object categories out of the box. Click "Auto-Annotate" → "Configure" to choose between your project model and the COCO base model.

Using a COCO base model

Select "Base YOLO model (COCO pretrained)" in the settings dialog. Map each of your project classes to the matching COCO category - e.g. your "Car" class maps to COCO "car". Only detections matching your mapped classes are kept; everything else is ignored.

Single-image auto-annotate

In the Annotator, navigate to an image and press A or click "Auto-Annotate" in the toolbar. Detected boxes appear instantly on the canvas - review and fix any mistakes.

Batch auto-annotate

Click "Batch Annotate" to run inference on all images in your dataset at once. A progress dialog shows the current file. When done, a summary shows how many images were annotated and any that failed.

Overwrite toggle

In the batch settings, the Overwrite option controls whether already-labelled images are re-annotated. Turn it off to only annotate images that have no labels yet.

Review and correct

Auto-annotation is never perfect. Use the unannotated filter to find missed images, and manually fix wrong or missing boxes. The goal is to reduce manual work, not replace it entirely.

Prepare your dataset

Make sure all your images are annotated. The Overview panel in the main window shows annotated vs total counts per class. Aim for at least 50–100 images per class for a basic model.

Split your dataset

Click "Split Dataset" in the tools panel. Drag the train/val/test sliders (default 70/20/10). Click Split - F-Annote copies files into train/, val/, and test/ subfolders inside your dataset.

Configure training

In the main window, set your epochs (50 is a good starting point), batch size (16 for most GPUs, lower if you get out-of-memory errors), image size (640 is standard), and select your base model. Give your run a name in the Name field - this is the folder name saved under runs/ for that training session.

Advanced settings (optional)

Click "Advanced Training" to access 30+ hyperparameters across 5 tabs: Model, Optimizer, Loss, Augmentation, and Control. The defaults work well for most cases.

Click "Train"

Click "Start Training" - the training window opens showing a live loss chart, epoch and batch progress bars, and a scrolling log. Training runs in the background - you can watch the loss curve converge in real time.

Training completes

When done, a results dialog shows your final mAP50 score and the path to your trained .pt file. The model is automatically added to your project's models/ folder.

What TensorRT does and when to use it

TensorRT is NVIDIA's inference optimizer. Converting your trained .pt model into a .engine file compiles it specifically for your GPU, giving 2-5x faster detection. It is completely optional - .pt models run fine everywhere - but if you plan to do live detection or batch auto-annotate large datasets, an engine is well worth the few minutes to build.

Step 1 - Confirm you have an NVIDIA GPU

TensorRT only works on NVIDIA hardware. Open Task Manager (Ctrl+Shift+Esc), go to the Performance tab, and check the GPU entry says NVIDIA - e.g. RTX 3060, RTX 4070, GTX 1660. AMD and Intel GPUs cannot build or run engines, so on those you simply stay on the .pt model.

Step 2 - Install the NVIDIA driver (the only download needed)

F-Annote already ships with TensorRT and the CUDA libraries bundled inside the app - there is no pip install, no CUDA Toolkit, and no separate TensorRT SDK download. The one thing you do need is an up-to-date NVIDIA GPU driver, which includes the CUDA runtime. Go to nvidia.com/Download, pick your GPU, download the latest Game Ready or Studio driver, install it, and reboot. That is the entire setup.

Step 3 - Verify CUDA is detected

Open F-Annote and check the Device dropdown in the training panel (or the Settings window). Your card should appear as "cuda:0 (Your GPU Name)". If you only see "cpu", the driver is missing or outdated - repeat step 2. When you try to build without CUDA, F-Annote also shows a clear message telling you what is wrong.

Step 4 - Build the engine

On the main window, pick your trained .pt model in the model dropdown, then click "Build .engine". F-Annote compiles the model for your exact GPU - this usually takes 1-5 minutes depending on model size and hardware. The app stays responsive while it works and notifies you when the .engine is ready.

Step 5 - Or auto-build right after training

Want to skip the manual click? Open Advanced Training, go to the Control tab, and enable "Pre-build TensorRT acceleration". The engine is then built automatically the instant training finishes, so your fastest model is ready with no extra step.

Step 6 - Use the engine

Once built, the .engine file shows up in the model dropdown right next to the .pt. Select it anywhere you run inference - single or batch auto-annotate, or Test Detection - and enjoy the 2-5x speed-up over the .pt.

Automatic cache and rebuilds

F-Annote stores engine metadata (GPU name, image size, model timestamp). If you retrain the model or switch to a different GPU, that cache is invalidated and the engine rebuilds itself automatically the next time it is needed - you never accidentally run a stale engine.

Open the Export dialog

Click "Export Dataset" in the tools panel. The export dialog opens.

Choose a format

Select one of the four formats: YOLO ZIP (for training elsewhere), COCO JSON (detectron2, MMDetection), Pascal VOC XML (legacy pipelines), or flat CSV (custom scripts and analysis).

Choose output location

Pick where to save the export. By default it goes to your project's exports/ folder. You can change this to any location.

Export

Click Export. A progress bar shows file-by-file progress. When done, a toast notification confirms success and shows the output path.

Click "Capture Screen" in the tools panel

This opens the screen capture tool - its purpose is to grab screenshots and add them directly to your project's dataset, not to run live detection. It is the fastest way to build a dataset from something visible on your screen.

Choose what to capture

Click "Select area" to draw a region on your screen with a crosshair snipping tool. Or click "Select window" to pick from a dropdown list of all currently running applications - e.g. your game, browser, or any other window.

Capture screenshots

Once a source is selected, press the capture button (or use the P global hotkey - it works even while your game or target app is in focus). Each press saves a screenshot directly into your project's images folder.

Add them to the project and annotate

Open the Annotator after capturing - your new screenshots are already in the image strip. Annotate them normally, then retrain to improve your model on real target data.

Click "Detect" in the tools panel

The Test Detection dialog opens. This is where you run your trained model on real images or a live feed to verify it is working correctly - separate from the screen capture tool.

Choose Screenshot or Live mode

Switch between the two tabs at the top. "Screenshot" mode runs inference on a single image capture. "Live" mode runs continuous real-time inference on a camera, screen area, or window feed.

Select a source

In Screenshot mode choose File, Camera, Area, or Window. In Live mode the same sources are available for continuous capture. "Window" drops a list of all running applications you can target.

Pick a model

Choose any .pt or .engine model from your project, or select one of the built-in COCO base models (yolov8n, yolov8s, etc.) to detect common objects without training. The label button next to the model shows "All" or the number of active class filters.

Filter classes (optional)

Next to the model dropdown is a button that says "All" by default. Click it to open the Detection Class Filter - a searchable list of all classes. Uncheck any classes you want to hide, use "None" to deselect all, or "All" to reset. Only checked classes will appear in results. The filter resets automatically if you switch models.

Run detection

Click "Detect". Bounding boxes with class labels and confidence scores appear over the image or live feed. Adjust the confidence slider to control how sensitive the detection is.