Perception is taking raw sensor data and creating different representations of the environment around the machine. This is one of Mojow's strategic areas of development. The most important output of the perception system is digital maps. They are used for all autonomous navigation tasks. There are many different approaches and methods for solving perception problems. Deep learning is one of them. We understand the importance of clean and organized data for data-driven applications.

• Our data repository has more than 100TB of targeted and organized datasets from a variety of sensors. This includes millions of data frames from a large variety of farming operations which feed Mojow's Data Engine for data-driven application development.
• Individual models are currently being developed for multiple specific farming applications.

Our machine learning pipeline supports various different operations including:

• Dataset annotation, management and tracking
• Training and retraining object detection and segmentation models
• Evaluation scripts for trained models
• Model testing scripts using our hand-picked and curated test sets
• Model deployment to edge devices
• Model tracking and management

Our dataset repostiory has more than 1,000,000 unique frames from different farms and from real operations. Our approach to dataset creation is very simple: no annotation errors should exist in our training data. To achieve this level of quality in our training datasets, we have a multi-level review process for our data annotation to guarantee no wrong labels/annotations make it to the training. Our training datasets come with segmentation annotations. Below are a few important annotation types that come with our datasets:

• Various types of boundaries such as crop boundary, tilled area, water boundary, road boundary, etc.
• Different kinds of rural roads: paved, gravel and dirt roads
• Farm field entrances and transition areas
• Objects of interest such as rocks, ruts, water, sloughs, foreign objects, etc.
• Farm equipment, implements and any other machines typically seen at farms
• And many other

Currently we're going through our millions and millions of collected datasets from various field operations. Thanks to our flexible machine learning pipeline, we're growing our annotated datasets for many more topics of interest. Our machine learning pipeline allows us to easily add support for new detection or segmentation tasks and prototype a solution to a new problem very quickly. This means that we can take any project from an idea in the office to field testing it in a matter of weeks and up to a few months.