This repository is meant to help you create Docker containers dedicated to image annotations. Currently, this is based on :
You must have docker installed.
Please note that if you installed and use docker desktop, it starts with a context called desktop-linux. In this context, you can have access to GPU only with sudo privilege.
If you don't want to run with privilege, you must run docker in rootless mode. In this case, you can switch to rootless context by running
docker context use rootless
If you want the containers to have access to the NVIDIA GPUs please follow the instructions given in docker : Access an NVIDIA GPU. Briefly, you will have to install NVIDIA driver and NVIDIA container toolkit. Once done, you should be able to successfully run the following command :
docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi
If you need sudo privilege for this command, please read the remark above (same paragraph).
If you run with rootless privilege, all the folders (volumes) that will be created with the containers (e.g. fiftyone_data, label_studio_data etc.) will NOT be owned by your current user. Which may be anoying for some usage. In this case, you may chmod -R g+w each folder.
Start docker.
Clone this repository with the following command :
git clone https://github.com/borisboc/docker_image_annotations.git && cd docker_image_annotations
Edit the file fiftyone_image/fiftyone_secrets.env. Change the values of the 2 environment variables :
MONGO_INITDB_ROOT_USERNAME=YourMongoDbUsernameHere
MONGO_INITDB_ROOT_PASSWORD=YourMongoDbPasswordHere
In order to set a proper user name and password for the MongoDB database (used by FiftyOne).
Build and start the containers with the following command :
- If you have GPU access :
./start_gpu.sh
- Otherwise, for CPU :
./start_cpu.sh
Open your web browser and and go to URL : http://localhost:8080/. Then log in label-studio.
Sign-up / create an account.
Get your label-studio API KEY (acces token) by going to the account & settings page. Copy the Access Token, and paste it inside the file label_studio_image/label_studio_secrets.env
LABELSTUDIO_API_KEY="PLEASE PASTE YOUR LABEL STUDIO ACCESS TOKEN HERE, WITHIN THE DOUBLE QUOTES"
Stop the containers (e.g. CTRL+C in the terminal you upped the containers, or using Docker Desktop).
Restart the containers (see docker compose command above) so that FiftyOne environment is updated.
Start docker.
Start the containers with the following commands :
Build and start the containers with the following command :
- If you have GPU access :
./start_gpu.sh
- Otherwise, for CPU :
./start_cpu.sh
Or restart it from docker desktop, as you prefere.
Create a subfolder inside folder local_images/ with the name of your project. E.g. local_images/myproject. Place your images within this folder.
Open your web browser and and go to URL : http://localhost:5151/. to use the FiftyOne App.
If you don't have any dataset yet, you can create one by clicking on link on the webpage. TODO : Screenshot when creating a dataset for the first time.
Then you can add samples by clicking on the link click here on the webpage.
Choose what you want to import. It depends on your data : if you have only media (i.e. the images) or you also have the annotations. For more details, you can refere to FiftyOne import_samples operator.
Then import either all the content of a directory (but not the subdirectories recursively). Or using a glob pattern.
In both cases, your images will be in the folder /home/local_images as there is a binding between the container and the folder on your host machine.
In the case you want to browse the folder with the FiftyOne Web App, see the screenshots bellow.
Access the root directory by clicking on the arrow up close to the text bar.
Then go to your directory containing your media. E.g., if your subfolder is named myproject (as proposed above), you will find it at the path /home/local_images/myproject
.
Then press execute.
From now you have a dataset created, and some samples. You can see the subparagraphs of this readme depending on what you want to do next.
And then request for annotations with Label-Studio backend.
TODO : finish this paragraph and add screenshots.
During annotations on Label-Studio, if you need to annotate segmentation (e.g. semantic segmentation, or instance segmentation) you can use Segment Anything Model 2 (SAM2) to generate you segmentations/masks from a provided seed on the image (which can be some keypoints, a brushed region, or a axis-aligend rectangle). The result of SAM2 will be BrushLabel.
You can find a youtube video tutorial from Label Studio, to show you how to use SAM2.
Concerning the configuration in the menu Settings > Model, in our case, please set the following Backend URL :
http://img-annotations-sa2-label-studio:9090
For further details, please refere to label-studio-ml-backend github : using sam2
TODO : finish this paragraph and add screenshots.
To start a jupyter notebook server on the label-studio container, simply start a new terminal and run :
./jupyter_label_studio.sh
You will see the URL with the AUTH token printed in the console. Go to this URL on the web browser to work with jupyter notebook. For instance, you may want to use the scripts that are in PythonUtils/ folder.
Same as for label-studio container, but instead you may run :
./jupyter_fiftyone.sh
You can use Visual Studio remote connexion on a Dev Container. Or Docker Desktop.
Connect on the container you want to play with (let's say img_annotations_fiftyone ).
Inside a container terminal, go to relevant folder. E.g. /home/local_images/ , then type
jupyter notebook --ip 0.0.0.0 --no-browser --allow-root
Click on the link provided in the terminal output to start a notebook.
label-studio : persist data with Docker
label-studio tuto import yolo annotations
label-studio-ml-backend github : using sam2
Access Jupyter notebook running on Docker container
FiftyOne : Loading data into FiftyOne
FiftyOne : Using the FiftyOne App
Screenshot when creating a dataset for the first time.
More documentation and more screenshots.