*3.2. Experimental Details*

We train ShadowDeNet based on the stochastic gradient descent (SGD) optimizer [75] by 12 epochs. The learning rate is set to 0.008, the momentum is set to 0.9, and the weight decay is set to 0.0001. The training warmup is the linear mode with 500 iterations. The learning rate is reduced by 10 times at the 8th and the 11th epoch. The training batch size is set to four because of the limited GPU memory. The input image size of ShadowDeNet is 660 pixel × 720 pixel. To avoid overfitting, the ImageNet pretrained weights [76] of ResNet-50 [50] are loaded for transfer learning. Other network parameters are initialized using the Kaiming method [77]. Other hyperparameters not mentioned are same as Faster R-CNN. During inference, the nonmaximum suppression (NMS) [78] is used to suppress duplicate detection boxes with an IOU threshold of 0.50.
