Skip to main content
. 2019 Oct 31;10:1404. doi: 10.3389/fpls.2019.01404

Table 3.

Comparison of different variants of Faster-RCNN and the proposed method on the paddy–millet dataset.

Method #parameters (million) Paddy Millet mIoU (%) Precision (%) (d_thresh = 15) Recall (%) (d_thresh = 15) F1 score
Paddy Millet Paddy Millet Paddy Millet
Faster-RCNN ∼136 50.07 46.37 48.22 95.42 94.69 74.87 68.58 83.90 79.54
EDNet ∼15.27 57.15 45.52 51.34 90.0 86.29 92.30 68.59 92.19 76.42
UNet ∼2.14 48.65 42.62 45.64 91.86 84.37 81.02 69.23 86.10 76.05
FCN8 ∼38.16 53.30 45.40 49.36 89.29 77.07 89.74 77.56 89.51 77.31
DeepLabV3 ∼4.14 15.93 43.27 29.61 51.58 95.69 33.33 57.05 40.49 71.48
ESNet (proposed) ∼5.74 56.53 47.02 51.78 87.80 84.56 92.30 80.76 89.99 82.16

EDNet, encoder–decoder network; ESNet, enhanced skip network; mIoU, mean intersection over union. The performance is quantified using intersection over union (IoU), precision, and recall. For Methods, bold is used to highlight the proposed method, whereas bold numbers are used to highlight the best results.