Tutorial 2: The MusicBox Detector

In this tutorial, we will demonstrate how to execute the four most important steps in every Machine Learning pipeline: Data Collection, Data Annotation, Model Training, and Model Deployment.

The process begins with the Maivin Platform. Using the web interface on the device, we can start the recording process. Once data is recorded, we can access the recording files and import them into EdgeFirst Studio. Studio will create a dataset from the recording. Once the dataset is created, we can begin the annotation process to produce both bounding boxes and segmentation masks. For this tutorial, we will focus solely on bounding boxes.

The duration of the annotation process depends on the dataset size. For this specific dataset, which contains approximately 150 images, the entire annotation process takes around 3 minutes. After annotation, we need to partition the dataset and use ModelPack for training. The training process will generate the model checkpoints, which will be deployed to the target device in the final step.

Data Collection

To begin recording a dataset, power on the Maivin platform and connect it to a network in order to access the Web User Interface (WebUI).

The default interface will load once you open the homepage on the target device in your browser (Google Chrome is recommended). For this tutorial, we’ll use only two out of the four options available on the WebUI:

• Segmentation Model: Displays the live camera feed along with the default people segmentation model.

• MCAP: Allows you to control recording options and provides download links for each recording stored on the Maivin device.

For a detailed walkthrough, see Web UI Walkthrough

The first recommended step is to go to the Segmentation view to check the camera’s field of view and ensure your target object is visible.

.

Once the object is clearly visible, return to the homepage and select the MCAP option to start recording. Click the Recording button (toggle on the left side).

The recording process takes a few seconds to initialize all services on the device. Once ready, the GUI will display the name of the recording (in red) that is about to begin. Now switch to the camera view and start interacting with the object — reposition it, and make creative or playful changes to the scene.

Reminder

Randomness helps improve dataset quality and enhances the model’s generalization ability.

Now it is time to download the dataset from the unit and upload it to Studio for annotation and model training. On the MCAP page, click the download icon next to each MCAP file and save it to your local PC.

Once stored locally, you’re ready to import the MCAPs into Studio via the Snapshot Dashboad.

In the GUI, click the FROM FILE button (top-right corner). This opens a multi-file selection dialog to upload all MCAP files at once. Studio will upload the files to the cloud for preprocessing. The GUI will show the name of one of the MCAPs as the default snapshot name. You can rename it by clicking on it (e.g., rename to MusicBoxTutorial).

Before restoring the dataset, you need to Create a Project and name it Tutorials

Restore Snapshot

Warning

Restoring a snapshot will incur costs against your EdgeFirst Studio account.

In the Snapshot interface, hover over MusicBoxTutorial (renamed above). A three-dot menu will appear on the right side of the GUI (next to "Available"). Click it and select Restore. See Restore Snapshots for details.

The GUI will prompt you to select a project and dataset name. Choose the Tutorials project and name the dataset MusicBoxTutorial.

Additional Options

The restore dialog provides advanced options like Topic Selection, Frame Rate, Depth Generation, and Automatic Ground Truth Generation. Leave all options at their default settings. These features are explained in AGTG

Click RESTORE SNAPSHOT. The dialog will close, and the dataset API will export the selected MCAP topics into a dataset, which will appear in the Dataset User Interface. Progress is shown in the Dataset UI.

The dataset is exported in EdgeFirst Studio Format

Data Annotation

Once the dataset is exported, we can see all the recorded sequences.

Once the dataset is exported, you will see all the recorded sequences.

In the dataset view, add the classes to be annotated. For now, we will add a single class: musicbox. Now we can start annotating the sequences. It's important that the dataset was imported as sequences, so the AGTG pipeline can use timing and tracking data to annotate most frames automatically.

Select any sequence to visualize it. Remember that MCAPs were exported at one frame per second.

If you are new to Studio, refer to Automatic Ground Truth Generation (AGTG) to learn how to annotate the dataset. Your dataset should now look like this in the Gallery (disable the Show Sequences option):

Dataset Partition

To train the model, it is essential to create training and validation groups within the dataset. The training set teaches the model, while the validation set is used to evaluate model performance during training.

On the dataset card, click the + button next to Group and create two groups: train and val. Assign 80% of the data to training and 20% to validation. If there are previous groups configured for the dataset, remove them prior to recreating the train and val groups.

The images are randomly distributed between both groups.

Data is randomly added to each dataset.

Model Training

As with other Studio features, Model Training has a dedicated user interface. See Model Training for details First, create an experiment to store all training sessions related to this dataset. Name it MusicBox Detection Tutorial. See Create Training Experiment for details.

Experiments are useful for comparing parameters across the same dataset and task, and for staying organized.

Inside the experiment, create a New Session and name it musicbox-detector. This name is also assigned to the cloud instance under the Studio console.

• Trainer: ModelPack

• Description: Object Detection

• Dataset: MusicBoxTutorial

• Groups: Select train and val

Most hyperparameters are auto-tuned by ModelPack, but some can be customized:

• Input Resolution: 640x360

• Model Name: Legacy (more model variants are going to be integrated in future versions)

• Epochs: 50 (default)

• Batch Size: Default is 16.

For small datasets, a large batch size may produce poor results. Use a batch size of 4 or 8.

Since our dataset is small, we need to use strong data augmentation to prevent overfitting and it is recommended to increase a bit the probability for each augmentation technique.

After creation, your session should look like this:

You’ll find general training info in this summarized view. Additional actions are available via the top buttons (top-right of the training card).

The training process begins with cloud instance initialization. Then the dataset is downloaded and cached. Training starts afterward. At the end of the training process, ModelPack quantizes the model and publishes the checkpoints.

Clicking the training card will show the expanded view containing all the logs and checkpoints.

From the available checkpoints, download the TFLite model - optimized for embedded devices — and deploy it to the Maivin unit for edge inference.

Uploading Models to the Raivin

In the SCP section in the SSH Tutorial, files can be uploaded to the Raivin using the command:

scp input_file torizon@verdin-imx8mp-XXXXXXX:.

Note

The above command assumes you have an SSH client, such as OpenSSH, installed. Please review the SSH documentation to confirm.

For the following examples, we will have the Fusion model of fusion.tflite and the ModelPack model modelpack.rtm that we want to upload to target device verdin-imx8mp-07130049.

First, we need to upload the files to the Raivin target using SCP:

$ scp fusion.tflite torizon@verdin-imx8mp-07130049:.
$ scp modelpack.rtm torizon@verdin-imx8mp-07130049:.

We should be able to confirm the files are there by using commands via SSH.

$ ssh torizon@verdin-imx8mp-07130049 ls fusion.rtm modelpack.rtm
fusion.tflite  modelpack.rtm

These files will be the /home/torizon directory, so their absolute filenames will be /home/torizon/fusion.tflite and /home/torizon/modelpack.rtm. Once the files have been uploaded, we can then configure the device with the files.

Deploying a New Model to the Model Service

There are two ways to deploy a new, 2D model to the Raivin's Model Service: the Web UI Interface or via the command-line.

From the Raivin Web UI

From the Model Service Configuration page, enter the absolute filename /home/torizon/modelpack.rtm in the "MODEL" text-box. Also, confirm the "Draw Boxes" check box is enabled, as the current trainer only supports 2D Box detection. Hit the "Save Configuration" box and continue on.

Remember to save the configurations at the end of the process. The Model Configuration page can be accessed via the following url: https://verdin-imx8mp-xxxxx/config/model

Manual Model Deployment

In case the manual deployment is needed, you need to connect to the device via SSH:

$ ssh torizon@verdin-imx8mp-15141030

and edit the model parameters in /etc/default/model

$ vi /etc/default/model

then restart the model service using the systemctl command

$ sudo systemctl stop model
$ sudo systemctl start model 

Note

Rember to use sudo to start and stop model services

Now the model is running, open the Raivin's WebUI, go to the Segmentation Page, and check the camera to see the model detection the object.

Begin testing the model with the object. If the model does not perform as expected, record a few more minutes of data and repeat the training process. Use this opportunity to identify edge cases and collect additional samples that can help improve the model's performance.

Conclusion

In this tutorial, we walked through the complete process of building and deploying an object detection model on an embedded device using the Maivin Platform, EdgeFirst Studio, and ModelPack. From data collection to model deployment, we covered few essential step of the machine learning pipeline:

• Data Collection using the Maivin Web Interface

• Data Annotation with EdgeFirst Studio

• Model Training with ModelPack

• Model Deployment on the Maivin unit

This workflow is not limited to object detection, the same process applies to any dataset type and to both segmentation and detection tasks, making it a powerful and consistent pipeline for embedded AI development.

Thanks to the Automatic Ground Truth Generation (AGTG) feature, the annotation process becomes significantly faster and easier—often requiring just a few clicks per sequence to annotate large volumes of data. This dramatically reduces manual labeling effort while maintaining high-quality ground truth data.

By following this tutorial, you now have a practical understanding of how to:

• Collect and prepare real-world data

• Use AGTG to automate annotation

• Train compact, optimized models with ModelPack

• Deploy models to the edge for real-time inference

Using this workflow ensures repeatability, scalability, and efficient development for embedded machine learning applications. Whether you're building a smart camera, an industrial monitor, or a self driving vehicle, or an edge AI prototype, this pipeline helps you go from raw data to deployment quickly and effectively.