How yolo9000 predicts big objects?

Question

I have a theoretical question about Yolo9000. Since it's a fully convolutional neural network, I was wondering how can it predict large objects. To the best of my knowledge, each convolutional kernel "knows" only local features, say 3x3, and Yolo9000 consists only of those. What trick does Yolo9000 use to detect object in all different sized anchors correctly?

Example:

Suppose I have big hi res image of a car, car area is about 10x20 of 32x32 standard yolo9000 anchors. According to the model central anchor should predict class "car" and its box sizes. How central anchor activation layer get information about side parts of such a car?

Answer 1

The receptive field of of a neuron in the final feature map can be larger than 3x3. Consider a simple neural network with 10 layers of 3x3 convolutions -- a given neuron in the last layer is affected by all the inputs in 21x21 area around it. Pooling also increases the size of the receptive field. Therefore, information from areas outside the immediate domain of the last convolution can be transmitted and used to make accurate detections.