> For AI agents: guidance on navigating Viam documentation is available at https://docs.viam.com/llms.txt. # Phase 1: Capture a starter set from your webcam Configure a webcam on your machine and add 30 images of dice to a starter dataset from the camera's control card. > Source: https://docs.viam.com/tutorials/dice-vision/capture-a-starter-set/ At the end of this phase, your webcam streams in the `camera` card's **Test** section and a dataset named `dice-starter` holds about 30 images of dice. You start small on purpose. In the next two phases you label these 30 images by hand and train a rough starter model on them, which then labels the rest of your images for you. ## Set up your workspace {#set-up-your-workspace} The camera doesn't need to point straight down, but keep it at the same height and angle for the whole workshop, except where [Phase 7](/tutorials/dice-vision/build-the-test-datasets/) asks you to move it. A model learns what dice look like from the distance and angle you train it on, so moving the camera changes what it sees. A steep angle, closer to straight down, makes the top face of each die easier to read, for you and the model. 1. Put a plain background to roll on, such as a dice tray or a sheet of plain paper, on a table with even light. 1. Mount the webcam about 30 cm above the background, aimed at it. 1. Plug the webcam into your computer. ![A webcam clipped to the top of a tall purple post on a white desk, angled down toward five dice on a sheet of white paper in front of the post, with the USB cable taped to the desk.](/tutorials/dice-vision/workspace-setup.jpg) ## Configure your webcam {#configure-your-webcam} The webcam discovery service lists every camera plugged into your computer and adds the one you pick as a camera component. That way you get your webcam, not a laptop's built-in camera. 1. In the [Viam app](https://app.viam.com), open your machine and go to the **CONFIGURE** tab. 1. Click **+** and select **Blocks**. 1. Search for `find-webcams`, select the **webcam-discovery** service, and add it to your machine. 1. Click **Save** and wait a moment for the service to start. 1. Open the discovery service's **Test** section. It lists the cameras connected to your computer. The same camera can appear more than once, once for each resolution and frame rate it supports. 1. Find your webcam and click **Add component** next to one of its entries. The new camera component is named after the device. 1. On the new camera component's card, click the **...** menu and select **Rename**. Enter `camera` as the new name. 1. Click **Save**. You can delete the discovery service from your machine once the camera is working; you don't need it for the rest of the workshop. > **Checkpoint:** > > On the **CONFIGURE** tab, the **Test** section of the `camera` card shows a live view of the background, and the dice are sharp. > If the image is from your laptop's own camera, delete the `camera` component and add a different entry from the discovery service's **Test** section. > If the dice are blurry, raise the camera a little until the webcam can focus. ## Create the starter dataset {#create-the-starter-dataset} A dataset is a named collection of images in your organization, which you label and train models on. The starter dataset is a small set of about 30 images that you label by hand in [Phase 2](/tutorials/dice-vision/label-the-starter-set/). In [Phase 3](/tutorials/dice-vision/train-a-starter-model/) you train a rough starter model on it, and in [Phase 4](/tutorials/dice-vision/auto-label-the-training-set/) that model labels the larger training set for you, so you only correct its boxes instead of drawing every one. The starter set lives in its own dataset because a training job rejects a dataset with more than 20% of its images unlabeled, and the training set stays mostly unlabeled until the starter model labels it. 1. Go to the **DATA** tab and click **DATASETS**. 1. Click **Create dataset**. 1. In the **Dataset Name** field, enter `dice-starter`. 1. Leave **Data type** set to **Binary Data**. 1. Click **Create dataset**. > **Checkpoint:** > > The **DATASETS** list shows `dice-starter` with 0 images. > If the name is taken, another dataset in your organization already uses it; pick a different name and use it wherever this workshop says `dice-starter`. ## Add 30 images from the camera {#add-30-images-from-the-camera} Each image you add here is one you'll label in [Phase 2](/tutorials/dice-vision/label-the-starter-set/), so aim for variety rather than volume. 1. On the **CONFIGURE** tab, open the **Test** section of the `camera` card. 1. Roll one to five dice onto the background. 1. Click **Add to dataset**, select `dice-starter`, and confirm. 1. Roll again and repeat until you've added 30 images. ![The camera card's TEST section showing a live GetImages feed of four dice on white paper, with the Add to dataset button below the image.](/tutorials/dice-vision/camera-test-card.png) Vary each roll: - Use different numbers of dice, from one to five. - Let some dice land near the edges and some in the middle. - Let two or three dice touch in a few images. - Make sure every face, 1 to 6, shows up at least 10 times across the 30 images. Turn dice by hand if a face is rare. A good starter set mixes how many dice are in each image, where they land, and whether they touch, like these three: ![Three overhead webcam images of dice on white paper: three dice spread apart, three dice near the right edge of the frame, and five dice with several touching in a row.](/tutorials/dice-vision/starter-examples.jpg) > **Checkpoint:** > > `dice-starter` on the **DATASETS** page shows 30 images, and each one shows dice from above. > If the count is lower, check that you selected `dice-starter` each time you added an image. In the next phase, you draw a box around every die in these 30 images and label it with its value.