Lobe Help

Everything you need to know to train great machine learning models with Lobe.

What is labeling?

What is labeling?

chevron-down

Labeling is assigning categories to your images to create examples that teach Lobe. These examples are commonly known as a dataset. From this dataset, Lobe will learn to automatically predict a label for a given image.


How do I create a dataset?

How do I create a dataset?

chevron-down

Label your images in Lobe.

  • Webcam - use any camera source to capture images directly in Lobe. You can optionally provide a label for these images. Hold down the camera button to capture a burst of images.
  • Image files - import common image files directly from your computer. Lobe supports JPEG, PNG, BMP, and WebP formats.

Import an existing dataset.

  • Folders - import existing labeled images by using folder names as the labels.

You can create new labels or edit existing ones by using the text box in the bottom corner of each image.


Note

  • The max image size Lobe can process is 178,956,970 pixels. For a square image, that is about 13,300 x 13,300 pixels.
  • When importing from an existing dataset, you can have a maximum of 4,994 images per label in a single import action. If you have more than 4,994 in a given label, please split up the dataset outside of Lobe so that each label has 4,994 or fewer images and import each split separately.


What types of images should I collect?

What types of images should I collect?

chevron-down

Collect images that you expect to see in the real world.

Lobe can only learn the patterns that exist in the images you provide as examples. Collect images from the same source you expect to use with your exported model.


Capture as many variations as possible.

Try to capture all the variations that naturally occur by collecting images in different conditions. Try different backgrounds, lighting, or zoom. This helps Lobe learn what parts of the image are useful for making predictions and what is noise.


Make sure your content is visible in the center square of the image.

While training your model, Lobe crops the center square of your images.


What labels should I use?

What labels should I use?

chevron-down

Label for each desired prediction.

Create a label for each category you want the model to predict.


Catch-all label.

Lobe will always predict one of your supplied labels for each image. If you expect the model to see images that don’t belong to any of your desired labels, create a None label as a catch-all bucket to show unrelated images.


How many images should I have?

How many images should I have?

chevron-down

Gather as many different images as you can.

A rough guideline is 100-1,000 images per label for smaller problems. To quickly get started with you project, Lobe has a minimum requirement of 5 images per label.


Have roughly equal number of images per label.

Balance the number of images between each label. If imbalanced, Lobe will be biased to predict the labels with more images and ignore the labels with fewer images.


Can I use videos?

Can I use videos?

chevron-down

You cannot use videos directly with Lobe, but it is a great way to collect many images. Turn your video into frames and select the best images for your problem.


You can use a tool like SnapMotion, VLC, Free Video to JPG Converter, or VirtualDub.


Can I connect external cameras?

Can I connect external cameras?

chevron-down

Yes! Lobe will use the default camera connected to your computer as the primary source. To switch cameras, use the source selector in the Label or Play views.


If you do not see an image from the camera, make sure the selected camera is connected to your computer and functional, and the Lobe app has permissions to use cameras. You can check permissions by going to:

  • Mac - System Preferences > Security & Privacy > Privacy > Camera
  • Windows - Settings > Privacy > App permissions > Camera


How do I speed up labeling?

How do I speed up labeling?

chevron-down

  • Use shift + click or cmd/ctrl + click to multi-select many images and label them at the same time.
  • Capture a burst of labeled images from the webcam.
  • Use folder names as labels to import existing datasets.
  • Use your arrow keys to move the image selection and label entirely with your keyboard.


How do I import images from a CSV?

How do I import images from a CSV?

chevron-down

You can’t use CSVs directly with Lobe. However, we built an external desktop app and Python command line tool to both download images and run your exported model on a CSV of image URLs.


Learn more on our GitHub.


Train your app with Lobe
Download