Create the Dataset

Vision custom models are intended for users without a data science background. By creating a dataset, and instructing Vision to train a model based on the dataset, you can have a custom model ready for your scenario.

Data labeling is the process of identifying properties of records, such as, documents, text, and images, and annotating them with labels to identify those properties. The caption of an image and identification of an object in an image are both examples of a data label. You can use Oracle Cloud Infrastructure Data Labeling to do the data labeling. For more information, see the Data Labeling service guide. Here is an outline of the steps to take:

1. Collect enough of images that match the distribution of the intended application.

   When choosing how many images are needed for your dataset, use as many images as you can in your training dataset. For each label to be detected, provide at least 10 images for the label. Ideally provide 50 or more images per label. The more images you provide the better the detection robustness and accuracy. Robustness is the ability to generalize to new conditions such as view angle or background.

2. Collect a few varieties of other images to capture different camera capture angles, lighting conditions, backgrounds, and others.

   Collect a dataset that's representative of the problem and space you intend to apply the trained model on. While data from other domains might work, a dataset generated from the same intended devices, environments, and conditions of use, outperforms any other.

   Provide enough perspectives for the images, as the model uses not only the annotations to learn what is correct, but also the background to learn what is wrong. For example, provide views from different sides of the object detected, with different lighting conditions, from different image capture devices, and so on.
3. Label all instances of the objects that occur in the sourced dataset.
   Keep the labels consistent. If you label many apples together as one apple, do so consistently in each image. Don't have space between the objects and the bounding box. The bounding boxes must closely match the objects labeled.

   Important
   
   Verify each of these annotations as they're important for the model's performance.