Image labeling or annotation is fundamental for every Artificial Intelligent (AI) project. The performance of the trained model is highly dependent on the quality of annotations because this process transfers the human “wisdom” into the AI domain. Lack of clear project objectives, annotation rules, and annotator’s experience are the most common reasons many AI projects fail to achieve their goals. This challenge is even more evident in the agritech domain, where image and object diversity, combined with limited agronomic knowledge from the annotator’s workforce, significantly compromises AI performance. Eden Library hosts thousands of high-quality annotated datasets and in this post, we attempt to share some valuable tips and rules to follow for optimal annotation in plant images.
Simply referring to how loose or tight a bounding box, polygon, etc., has been drawn around the object. In agriculture, instances like leaf diseases are hard to distinguish, so annotations need to be spatially accurate and include only the relevant visual features. Nonetheless, it’s a matter of “what” a model should learn, so if your application requires the detection of entire diseased leaves and not just individual symptoms, then the annotation should be as tight as possible on the leaf level.
Another critical aspect to consider is the annotation consistency or homogeneity. You cannot afford to have objects that, in some cases are loosely annotated and in other cases are tight. This will ingest features in the training process that are irrelevant and out of your project’s scope. Consistency is also important in agriculture cases where plants, symptoms, or instances that are difficult to identify, are misrepresented, or noisy. In such cases, you need to establish clear rules about “what” you annotate and “how” tolerant you can be. Ensure that you respect those rules throughout your entire dataset.
It might sound like an obvious rule, but in agriculture use-cases, where you can find a few hundred objects in a single image, some instances are not annotated. Modern computer vision architectures can be robust enough to “learn” the visual patterns even if you have not annotated 100% of the objects. Still, if your dataset is limited and the objects are visually diverse, this can be an extra nail in your project’s coffin.