Introduction
In this tutorial, you'll learn how to use Voxel51 3.0, a powerful platform designed for physical AI teams to manage multimodal sensor data. This platform combines search, curation, annotation, and evaluation tools in one place. We'll walk through setting up a basic project using Voxel51's self-service tier and demonstrate how to upload and organize your multimodal data.
Prerequisites
- A basic understanding of what multimodal data means (data from multiple sources like images, videos, and sensor readings)
- Access to the Voxel51 3.0 platform (you can sign up for the free tier at voxel51.com)
- Sample multimodal data files (images, videos, or sensor logs) to work with
- A web browser (Chrome, Firefox, or Edge recommended)
Step-by-Step Instructions
1. Create Your Voxel51 Account
First, visit the Voxel51 website and sign up for an account. The free tier allows you to explore the platform's features without any cost. You'll need to provide your email and create a password.
2. Set Up Your First Project
After logging in, you'll see the dashboard. Click on "Create New Project" to start a new multimodal data project. Give your project a descriptive name like "Physical AI Sensor Data Analysis" and select the appropriate project type. This step creates a container for all your data and annotations.
3. Prepare Your Multimodal Data
Organize your data files into a single folder structure. For example, you might have:
- Images folder with .jpg files
- Videos folder with .mp4 files
- Sensor logs folder with .csv files
This organization will make it easier to upload and search later. Each file type represents a different modality of data that physical AI systems often work with.
4. Upload Your Data to Voxel51
Once your project is created, navigate to the "Data" tab. Click "Upload Data" and select your organized folder structure. Voxel51 will automatically detect the different file types and organize them accordingly. This process is crucial because it creates a searchable database of all your multimodal data.
5. Explore the Dataset Versioning Feature
After uploading, you'll notice that Voxel51 automatically creates a version of your dataset. This is important for physical AI work because sensor data often changes over time. You can create new versions as you collect more data or refine your datasets. Click on "Versions" to see how your data has evolved.
6. Use Voxel51 Agent Tools
Voxel51 3.0 comes with 45+ MCP tools through its Agent feature. These tools help with data annotation, evaluation, and search. For example, you can use the "Object Detection" tool to automatically identify objects in your images or videos. Navigate to the "Agent" tab and explore the available tools. Each tool is designed to work with different aspects of multimodal data processing.
7. Annotate Your Data
Click on "Annotation" in the sidebar to start labeling your data. You can draw bounding boxes around objects, label different sensor readings, or mark specific events in your videos. This annotation process is essential for training physical AI models because it provides the ground truth data needed for machine learning.
8. Evaluate Your Dataset
Go to the "Evaluation" section to assess the quality of your data. You can use built-in tools to check for consistency, completeness, and other quality metrics. This step ensures that your multimodal dataset is reliable for physical AI applications.
9. Search and Curate Your Data
Use the search functionality to quickly find specific data points. You can search by tags, file types, or even content within images. This search feature is particularly powerful for physical AI because it allows teams to quickly locate relevant data for specific use cases or experiments.
10. Export Your Results
When you're ready to use your data outside of Voxel51, you can export your annotated datasets. Click on "Export" and choose the format that works best for your downstream AI applications. This could be standard formats like COCO for object detection or custom formats for your specific physical AI models.
Summary
In this tutorial, you've learned how to set up a Voxel51 3.0 project, upload and organize multimodal sensor data, and use the platform's powerful tools for annotation and evaluation. The platform's unified approach to managing different types of physical AI data makes it an excellent choice for teams working with complex sensor datasets. By following these steps, you've created a foundation for building physical AI applications that can leverage multimodal data effectively.
Remember, the key to success with Voxel51 3.0 is to start simple and gradually explore more advanced features like dataset versioning and the extensive MCP tools. The platform's free tier is perfect for getting started and testing different workflows before committing to paid features.
