Fig. 8: Training the detection and segmentation models on the synthetic and real biofilm datasets.
From: Deep generative modeling of annotated bacterial biofilm images
a, b Comparison of Mask R-CNN models trained on two different synthetic datasets: one with random cell locations and one with specific patterns. The second model showed better results in the detection and segmentation of paired cells, and these results can be observed both visually and quantitatively. c Comparison of two Mask R-CNN models on a real biofilm dataset. d The relationship between the IoU score and the number of images in the training dataset. The IoU reaches a plateau at 110 images. e Comparison of Mask R-CNN trained on the real manually annotated dataset with two models trained on generated datasets. Gray lines denote percentile intervals.
