Abstract: Implementations are described herein for improving unsupervised domain adaptation (UDA) by using improved adaptive teacher for object detection with cross-domain mix-up. In various implementations, cross-domain training of an object detection machine learning model may include: performing weak augmentation on images from a target domain DT to generate a first set of weakly augmented target domain images; perform strong augmentation on images from the source domain DS and images from the target domain DT to generate a second set of strongly augmented images; processing the second set of strongly augmented images to generate a third set of inter-domain mixes of the images from DS and DT; and jointly train the object detection machine learning model, as a student machine learning model, with a teacher machine learning model using the first and third sets.

Abstract: Implementations are described herein for learning mappings between a domain specific language (DSL) and images, and leveraging those mappings for various purposes. In various implementations, a method for using a DSL to generate training data may include processing data indicative of ground truth image(s) depicting a real plant using a trained image-to-DSL machine learning (ML) model to generate a first expression in the DSL that describes structure of the real plant. The first expression may include a plurality of parameters, and may be processed to programmatically generate a plurality of synthetic DSL expressions. Each respective synthetic DSL expression may describe structure of a respective synthetic plant for which parameter(s) have been altered from the first expression. The synthetic DSL expressions may be processed using a renderer to create three-dimensional (3D) synthetic plant models. Two-dimensional (2D) synthetic images may be generated that depict the 3D synthetic plant models in an area.

Abstract: Implementations are described herein for reducing the time and costs associated with the crop scouting in a crop field. In various implementations, a method is implemented using one or more processors, and the method include: operating, based on a type and arrangement of a crop field, a robot to travel along a trajectory through the crop field using a first gait. The robot includes one or more vision sensors. The first gait includes a first repeating cycle of poses of the robot. The method can further include: synchronizing operation of one or more of the vision sensors with one or more poses of the first repeating cycle of poses of the multi-legged robot to capture one or more initial sequences of images depicting one or more points-of-interest of crops growing in the crop field.

Abstract: Implementations are disclosed for dynamically allocating aspects of platform-independent machine-learning based agricultural state machines among edge and cloud computing resources. In various implementations, a GUI may include a working canvas on which graphical elements corresponding to platform-independent logical routines are manipulable to define a platform-independent agricultural state machine. Some of the platform-independent logical routines may include logical operations that process agricultural data using phenotyping machine learning model(s). Edge computing resource(s) available to a user for which the agricultural state machine is to be implemented may be identified. Constraint(s) imposed by the user on implementation of the agricultural state machine may be ascertained.

Abstract: Implementations are described herein for performing depth estimation in the agricultural domain, including generating synthetic training data. In various implementations, one or more three-dimensional synthetic plants may be generated in in a three-dimensional space, wherein the one or more three-dimensional synthetic plants include homogenous and densely-distributed synthetic plant parts. The plurality of three-dimensional synthetic plants may be projected onto two-dimensional planes from first and second perspectives in the three-dimensional space to form a pair of synthetic stereoscopic images. The first and second synthetic stereoscopic images of the pair may be annotated to create a mapping between the individual synthetic plant parts across the first synthetic stereoscopic images. A feature matching machine learning model may be trained based on the mapping.