Upload and Tag Images Multi-Label
The next step is to upload our training data and begin labelling it with multiple tags per image.
1. Upload Training Images
Click Add images. Start with 10 images from the ai_services/fruitbowl_dataset/Training folder on the scryptIQ GitHub repository. These images contain examples of all fruit types, in different patterns and combinations.
Training images showing various fruit combinations in baskets - each image can contain multiple types of fruit that need to be tagged individually
Important
Custom Vision requires a minimum of 5 tagged images per class in order to begin training. It will not train if this condition is not met. The more training images and tags you provide, the more robust the model's performance will be. Our sample dataset includes exactly 5 instances of each fruit type, across all 10 images.
2. Tag the Images
Tag the Images manually using the interface. You can apply multiple tags per image. The more accurately and consistently you tag, the better your model will perform.
Custom Vision interface showing how to apply multiple labels to a single image containing various fruits
Tip
Include background-only images (such as an empty basket) in your training set and explicitly tag them with a background label (such as 'basket' or 'empty'). This teaches the model to recognise when target objects are genuinely absent, significantly reducing false positive predictions. Without these negative examples, the model may incorrectly associate background textures, colours, or shapes with your target classes. We'll explore this a little further, shortly.