Not understanding the inputs to the pretrained model

[91parthparikh]

During inference part we send two inputs to the network.

output,feature = net(input,inputcoord)

As input is a patch size of 96x96x96 and inputcoord is a patch size of 24x24x24x3. When I looked at the values of inputcoord they are ranging from the -0.5 to 0.5. I am unable to understand the reason for having a negative valued anchor boxes. Generally faster rcnn has normalized anchor boxes of image size with set of height and width, where as here all the values for inputcoord are ranging from [-0.5, 0.5].

Is there a relation between input and inputcoord ? Because input and inputcoord are normalized in [-0.5, 0.5] range.

Is it a different way to initialize the anchor boxes ? It would be helpful if you can explain it briefly.