Apr 22 Image Fig. 1. Demonstration of the bottom-up method for root pose estimation. A detailed image of a root system cultivated in a clear gel cylinder is input into a trained NN via SLEAP. This network concurrently produces confidence maps and part affinity fields. The confidence maps, with distinct color codes for each landmark type (r1 in pink, r2 in yellow, r3 in green, r4 in purple), pinpoint probable root landmark positions. These are then sharpened into discernible peaks based on the highest probabilities within each map. Part affinity fields encode part-to-part associations in 2D vector fields across the image. Using these data, connections between root landmarks are evaluated and scored. The most highly scored connections serve to pair the root landmarks, culminating in the assembly of individual root instances, each distinguished by a unique color. Elizabeth Berrigan SPLS Seminar: Automated Root System Phenotyping Using Deep Learning 4 – 5 p.m., April 22, 2025