iDS NXT Ocean: A simplified Artificial Intelligence workflow experience

Following this, users will select the data set they wish to use. The data set tab, located near the top of the page, provides your chosen thumbnail image, a summary of the data set, a number of corresponding annotations and the number of images chosen for the data set. On the right of the data set tab you can select the utilization drop down box, which allows the selection of what proportion of images will be used for training and what percentage will be used for validation. The default is a 70/30 (training/validation) split but users can customize based on their own preference.Once you have selected your data set and what proportion of the images will be used for training and validation you can test your data set for training.

The first part of the training section provides the opportunity for data augmentation - with options to rotate, zoom, shift, distort and shift colour channels. This is done to help improve thegeneralization of the neural network. Users can use the slider to control the extent of the transformation. Flipping images is only relevant for applications where the object does not have a logical orientation. For example, mirroring text would not be beneficial to for your intended application. In contrast, with objects like fruit and nuts if all the available transformations are applied, the outcome for your intended application would typically be improved.

The next section is related to the neural network requirements. In this section, users can choose the iDS NXT OS version - in most cases it is recommended to use the latest version, however for compatibility with existing deployed systems the appropriate version can be selected from the drop down list. Users will now choose the inference time. You can choose options between 13-milliseconds to 75-milliseconds - the appropriate selection is based on the necessary decision time for the system balanced against the accuracy requirements. In some cases it can be helpful to create multiple neural networks with different inference times in order to compare performance. Once you have selected these parameters click "train neural network" to begin the training process.

During the training NXT Lighthouse provides a graph which documents the achieved accuracy for the training sequences of the neural network.

The above example is for classification and the process of training neural networks for object detection is quite similar. Users follow the same process but at the start of the project creation we select detection as the project type. Similarly, users enter the project name, description and add a thumbnail and email addresses for the notifications.