FarmScapes - the foundation of EYEBOX™

Every deep learning model is only as good as what it's been shown. Long before EYEBOX™ ever makes a decision in a field, such as stopping for a rock pile, adjusting a boom for a dip in the terrain, or holding a line through dust, it has learned to recognize that situation from data. In agriculture, that data is brutally hard to get right. Fields change by the hour, let alone the season. Dust, glare, mud, stubble, shadow, and rain all distort what a camera sees. Most perception models built for structured environments simply weren't trained for this kind of chaos.

That's the problem FarmScapes was built to solve. It's Mojow's proprietary agricultural dataset and it has become the foundation underneath everything EYEBOX™ does in the field.

500TB+

OF CURATED

DATA

100M+

FRAMES

5 Seasons

2020-2025

150+

METADATA

LABELS

A dataset built at farm scale

FarmScapes consists of hundreds of millions of frames and more than half a petabyte of curated agricultural data. This data is collected the only way that actually produces reliable results - on real farms.

The collection spans five consecutive farming seasons, from 2020 through 2025, across Alberta and Saskatchewan. It covers every major stage of the farming calendar: seeding, spraying, and harvesting. It also covers every condition a machine will actually encounter: daylight, dawn, dusk, darkness, clear skies or rainy days, as well as the kind of dust that a wheat field kicks up in August. All of it was gathered from real farm operations, with owner permission, instead of staged or simulated conditions.

A perception model that has only ever seen a clear, sunny field will struggle the first time it meets a shadowed headland at dusk, that is why the range of conditions matter. FarmScapes is built so that struggle happens in the training stage and not in a customer's field.

Sensor-rich and precisely synchronized

Every scene in FarmScapes is captured through a full sensor suite. That includes ten or more high-resolution RGB cameras in a surround configuration, LiDAR for 3D point cloud, and RTK GNSS paired with a precision IMU for six-degrees-of-freedom positioning. All of it is synchronized to sub-millisecond precision, with intrinsic and extrinsic parameters fully calibrated across sensors.

In practice, that means every frame isn't just an image, but a precisely time-aligned, spatially calibrated snapshot of a real working environment, that is ready to support object detection, semantic segmentation, and depth estimation models alike.

Annotation you can actually trust

Raw sensor data is only half the story. The other half, which is arguably the harder half, is annotation: labeling what's actually in each frame, consistently and accurately, at scale. Mojow made a deliberate choice here - we don't outsource it.

Our annotation pipeline is 100% in-house, run by trained internal annotators and reviewers rather than a rotating pool of third-party contractors. Every dataset carries over 150 metadata labels, enabling fine-grained retrieval for whatever a model needs, whether it's a specific crop stage, lighting condition, obstacle type, or terrain feature. Every annotation passes through a two-tier, independent review process before it's considered training-ready.

It's a slower, more deliberate way to build a dataset. It's also the only way we're willing to build one. Annotation quality determines model quality, and model quality is what stands between an autonomous machine and a bad decision in the field.

Why this matters for OEMs

For OEMs that are building perception models for autonomous equipment, FarmScapes removes one of the largest hidden costs in the process: sourcing and validating training data good enough to trust. Rather than assembling a patchwork of limited datasets or depending on external annotation vendors with inconsistent quality and expertise, OEMs can build on data that has already been curated, cleaned, and reviewed to be compatible with any deep learning architecture.

It's the same principle behind everything Mojow builds - don't ask the industry to start from scratch. FarmScapes gives OEMs and developers a running start, real-world scale and diversity, precise sensor calibration, and annotation quality that has been controlled end-to-end.

Built on real farms for what comes next

FarmScapes exists because Mojow's own technology needed it first. Every season of data in this collection has helped train and refine the perception behind EYEBOX™, and it continues to grow with every season that we're in the field. That is the advantage of building a dataset the way we build everything else at Mojow - on real farms, with real farmers, and with an eye toward what autonomous agriculture actually needs.

To learn more about FarmScapes and how it can support your perception models, visit mojow.ai/datasets.