I m confused of how one whole bounding box is decided? In m real life very small grids can a part of car and that whole grid is bounding box w.r.t to grid. bx,by,bh and bw are decide based on the grid boundries. So how this grids are combined to make one bounding box ?
Maybe take a look at this existing thread, and let us know if it helps?
No I m still confused about how you predict one whole boundary out of multiple grids detecting the same object.
Also in above link, are anchor sizes have fixed height and width value or they are being generated based on particular grid boundaries?
Anchor sizes are fixed. YOLO uses k-means unsupervised learning to select a set. Detailed explanation here
Each neural net output (grid cell location + anchor box) makes its own predictions. Each prediction includes one whole boundary box. There is no sharing or merging or combining.