Seed images are the core input to the Dataset Generator. They function like stickers that will be pasted onto different backgrounds to generate large amounts of data for model training. Each seed image must be a transparent PNG, where the background has been masked out, leaving only the desired object visible.
What makes a good seed image?
Good seed images are representative of the real-world data we are simulating. They should be as diverse as possible and contain as many variations of the subject as you expect to see in real world use. For example, if you are training a pose estimation model to detect the position of fingertips, your seed images should contain hands of different genders, ages, and skin tones. Seed images should also capture subjects in a variety of positions and lighting conditions.
If your object typically appears in a context with other objects, they should also be included. You will be given the opportunity to annotate seed images to guide models to predict only certain elements of your subject and ignore others. For example, if you were training an object detection model to detect wheels on a car, some of your seed images should feature the entire car, not just the wheel. This will help improve model accuracy by associating common context with your subject.
Good seed images are also large enough to composited at different scales. We recommend seed images measuring at least 512x512. If the object you are are trying to detect with your model is typically small in size relative to the image, smaller seed images can be used. E.g. identifying a basketball in a wide short of the whole court.
Creating seed images
There are a number of tools you can use to create seed images. Here are a few of our favorites.
remove.bg is a free service that utilizes server-side AI models to remove the background of images. To use it:
- Navigate to remove.bg,
- Select the image you want to remove the background from.
Downloadto save the result.
2. Preview (MacOS Only)
If you are using a mac, you can remove the background of an image with the built-in Preview app. To use this method:
- Open your image in Preview.
- Select the
Show Markup Toolbar.
- Select the Instant Alpha tool in the toolbar.
- Click and hold a pixel in the background of your image and drag the mouse away from your initial click. You will see the a red mask appear showing which parts of the image will be removed.
- Release the mouse when you have an area you want to remove. A selection will appear.
- Press the
backspacekey to delete the parts of the image inside the selection. If prompted, confirm image conversion to PNG.
- Save your image as a PNG.
If you have access to Photoshop, there are a number of methods for removing backgrounds depending on which version you have.
- Using the ML-powered Background Remover
- Using the Quick Selection Tool
- Using the Backround Eraser Tool
Microsoft Office 365 users can remove backgrounds from images on any platform by following this official guide.
A hand pose estimation model
Goal: Train a pose estimation model to predict the location of finger tips in live video.
Seed images feature a variety of hands from a diverse set of subjects. Even though the model will only predict the location of finger tips, hands also contain palms and wrists in most cases.
Goal: Train an object detection model to detect soccer balls in images.
Seed images are individual soccer balls with the backgrounds removed. A variety of soccer ball designs are provided.
Goal: Train an image segmentation model to detect face masks
Seed images are a mix of portraits of individuals wearing masks and masks themselves, not pictured on a person's face.