*4.2. Qualitative Results*

Figure 18 shows the qualitative moving target shadow detection results of different methods. In Figure 18, the display confidence threshold of Faster R-CNN, FPN, YOLOv3, RetinaNet, and ShadowDeNet is 0.50. However, CenterNet does not offer the similar classification confidence scores. In Figure 18f, the numbers above boxes denote CenterNet's Gaussian heatmap probabilities of the top five keypoints. Here, we exhibit the shadow detection results of the 601st, 772nd, 806th, and 900th frames.

*Cont*.

**Figure 18.**

(**g**)

**Figure 18.** Qualitative video SAR moving target shadow detection results of different methods. (**a**) Ground truth; (**b**) Faster R-CNN; (**c**) FPN; (**d**) YOLOv3; (**e**) RetinaNet; (**f**) CenterNet; (**g**) ShadowDeNet. The false alarms are marked by orange boxes. The missed detections are marked by red ellipses. Apart from CenterNet, the numbers above boxes are the confidences. The numbers above boxes in (**f**) denote CenterNet's Gaussian heatmap probabilities of the top five keypoints. The IOU thresholdis0.50,thesameasthePASCALVOCcriterion[79].

From Figure 18, one can draw the following conclusions:

