How YOLO training and prediction works for an object fall in multiple grid?

Question

So far what I understood about YOLO, it expects training image should be divided in to fixed grid, where each grid has Label like P(object present or not), object bounding box, object classes. Similarly it will return the same output for each image prediction.

If it's correct, I'm not able to map those images for both cases training and prediction where some objects are the part of multiple grids. During training we provide bounding box information corresponding to particular(single) grid only, how it clubbed the bounding box info of multiple grids?

Note: Non-max suppression is again confusing, if it is related with it.

Answer 1

It seems after referring many documents, I found the answer of my question.

First, likely to correct my understanding. I thought for labeling, bounding box size (width, height) will always be with in the range of particular grid dimension i.e. between 0 and 1. It's not correct and this is the only source of confusion atleast for me.

I believe that if I'm able to answer about the labeling process, I think it will clear most of the doubt raised in the form of question.

Steps are labeling:

1. Identify the center point of the image with respect to image boundary, refer RED dot.

2. Create a grid on top of the image.

3. Identify the grid where center point exist. In this case its 8th grid (Numbering start from 1).
4. Marked only that grid has an object(not others) and labeling should be something like this:

g8 is having object and bounding box information. bx and by are wrt to grid dimension, whereas width and height are wrt to its size, which is greater than one.

During YOLO model training, it learned this intelligence and can start predicting the height and width which can be greater than 1. In a case if a car( or any object) falls in multiple gird and model training is perfect it should predict the height and width accordingly i.e > 1.