TensorFlow object_detection model to identify person, for the same image, different image sizes have different recognition results,why?

I'm using ssd_mobilenet_v1_0.75_depth_coco from the TensorFlow 1 Detection Model(COCO-trained models) to identify people, and am now running into a strange problem. When the input image size is 2160x1080, the recognition result is null, but after scaling the image to 1920x1080, it can recognize the person correctly, does the model have any rule for the input image size? Is this a bug? What should be done about it?