Question about YOLOv1 bounding box coordinates

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

I’m struggling to create a light YoloV1 (with only one bounding box) on MNIST dataset (I randomly paste 28x28 digit into a 75x75 black background).

I can’t figure out how to turn relative-to-cell coordinates into absolute coordinates.

Since now, I’m using the groundtruth bounding boxes to retrieve the cell which should contain an object, then I save the i,j positions, then I use those positions to get back to absolute coordinates with my predictions.

This method works but when it’s time to detect a real image, I won’t have the groundtruth coordinates and so, the i,j object position, and so the absolute position of the predicted bounding box.

Could someone help me ?

2

The output of YOLO v1 has the shape S × S × (B ∗ 5 + C) where S is the number of grid cells. You can infer the i and j coordinates from the position in that output. As an example, the center cell of a 19x19 grid has i=8 and j=8. Since you know the cell indices, and you know how big each grid cell is in pixels, you can compute an image-relative pixel offset given a grid cell-relative offset.