Figure 6.

Test showing the focus of the network using a sliding window. (A) is the original test image, along with the count ground truth. Overlaid is a black sliding window (60 × 60 pixels), which traverses the original images. The top row is an example of a wild type plant and the bottom row is an example of a mutant. (B) represents the prediction accuracy as the sliding window is traversing the image of the multitask learning (MTL) model (left column) and the single task model (right column) trained with just 25% count labels. The errors are expected to be confined to the area where the plant is located as the box obscures whole or parts of leaves the overall count prediction should decrease. (C, D) represent the sliding box test only in the MTL model for Genotype classification accuracy and projected leaf area (PLA) estimation. These were only performed on the MTL model comparing between models trained with 100% (right column) and 25% (left column) labels. The color bar for the PLA task shows increments in percentage points. The rows correspond to the images in part A. GT signifies the ground truth.