*5.3. Ablation Study on SDAL*

Table 8 shows the quantitative detection results of ShadowDeNet with and without SDAL. From Table 8, SDAL can improve the overall performance of ShadowDeNet by ~3.8% *f* 1 accuracy, which shows its effectiveness. This is because SDAL can model the geometric deformations of the moving target shadow to adapt to motion speed variations. Finally, ShadowDeNet can discriminate the static backgrounds and the moving target shadows more effectively. Of course, SDAL is not free, and it usually needs more time to learn the kernel offsets of Equation (7) in training. However, once the kernel offsets have been learned, the speed sacrifice in the inference or test process is insignificant.

**Table 8.** Quantitative results of ShadowDeNet with and without SDAL.

