Figure 2.
Summary comparison of object detection algorithms for cellular detection. (A) Comparison of AP for all datasets showing that additional vertical data flipping is effective for raising accuracy in general. Without (TR1) and with vertically flipped data augmentation (TR2) and when trained using multiple datasets without (TR3) and with (TR4) vertically flipped data. The Friedman’s test was applied using Dunn’s multiple comparisons to compare TR1/TR2 (n = 24, AP ± SD, * P < 0.05) and TR1/TR3 (n = 24, AP ± SD, nonsignificant). (B) Overall average AP comparison of Faster-RCNN, YOLOv2, YOLOv3, and RetinaNet for each dataset, averaging across each of the training modalities (TR1–TR4), described in A. Numerically, the best-performing algorithm is marked with an arrow for each dataset case YOLOv2 (3/6) and RetinaNet (3/6; n = 4, AP ± SD).

Summary comparison of object detection algorithms for cellular detection. (A) Comparison of AP for all datasets showing that additional vertical data flipping is effective for raising accuracy in general. Without (TR1) and with vertically flipped data augmentation (TR2) and when trained using multiple datasets without (TR3) and with (TR4) vertically flipped data. The Friedman’s test was applied using Dunn’s multiple comparisons to compare TR1/TR2 (n = 24, AP ± SD, * P < 0.05) and TR1/TR3 (n = 24, AP ± SD, nonsignificant). (B) Overall average AP comparison of Faster-RCNN, YOLOv2, YOLOv3, and RetinaNet for each dataset, averaging across each of the training modalities (TR1–TR4), described in A. Numerically, the best-performing algorithm is marked with an arrow for each dataset case YOLOv2 (3/6) and RetinaNet (3/6; n = 4, AP ± SD).

This Feature Is Available To Subscribers Only

or Create an Account

Close Modal
Close Modal