How is partial input inferenced?

[kongyanye]

https://github.com/zoranzhao/DeepThings/blob/57c61b66abe1159ee7da16477dfac1b5b26800d8/src/inference_engine_helper.c#L323

I'm confused with how partial input/image is inferenced with the model. In the Line 323 above, the cropped_output of layer i-1 is directly fed as input for layer i for inference. But the net is actually the complete model. It seems each partition of the image flows into a whole model for inference while only the specific part in output map is fetched as results. But in that case, the input should be correctly placed in the same location as original image, instead of just simply feeding as inputs. Am I understanding correctly?

Anyone know it please help me!