SI Figure 3. Six images were randomly selected from the test set during
training that evaluate the performance of training on the final step
(50,000). On each image, the left side is the computer-generated image
and the right side is the human labeled image.
Appendix 3: Intersection over
union
The model was evaluated throughout the training process using
intersection over union (IOU), the degree of overlap between human
labeled and computer-generated identifications. Higher IOU represents a
greater overlap of the two. For our model, IOU did not depend on the
number of images input for training; rather, the uniqueness of objects
due to shape and texture was the determining factor. IOU graphs for all
object classes are displayed in Supplementary material Fig. 4.