Course4 Week3: Understanding YOLO Algorithm

[arunnalpet]

Hi,
I have several questions about how YOLO algorithm works. Here are my questions:

1. In the course Andrew mentions that Image classification is applied to each of the individual “grid cells”. In case where an object could span across multiple grid cells, how could the algorithm achieve to still identify the center point of the object, and span the bounding boxes beyond a particular grid cell?? What actually happens when object is spread across multiple grid cells?

2. In the forward pass, there are chances that algorithm might detect multiple bounding boxes for a single object. Why does this happen? The forward pass happens only once per image, so how are we ending up with multiple predictions of bounding boxes per image?? I know it finally gets suppressed, but how does it end up with so many bounding boxes in the intermediate stages?

3. For YOLO training, can we feed images with multiple objects in a single image? or should it be a cropped image with single class? How does the training image annotated? will it be 3x3x2x8? or would it be a 1x8 vector representing a specific class?

[ai_curious]

Maybe take a look at this existing thread and let us know if it answers any of your questions?

Detecting Multiple Objects using YOLO - Grid Cells plus Anchor Boxes

Detecting Multiple Objects in YOLO – Grid Cells plus Anchor Boxes One key feature of YOLO is its ability to detect (locate plus classify) multiple objects per image. Previous threads have introduced the concept of anchor boxes and how they are applied during training data setup. This thread will cover how grid cells and anchor boxes together enable predictions on multiple objects within one image. The motivation Think back to your first efforts to train a classifier for image based on a singl…

[arunnalpet]

Thanks for pointing out to detailed description!
Few things are clear, but still not able to connect the dots all together!

Like the earlier sliding window concept was very clear! But after that when we talk about YOLO, it is not clear on how each grid can detect an object which is larger than the grid. (However the representation of Grid and its corresponding outputs are clear).

Also, on my Question No.3, What would be the format of a single labelled training image?

[ai_curious]

@arunnalpet

Also, on my Question No.3, What would be the format of a single labelled training image?

[How to prepare bounding box labels - #5 by ai_curious]

[ai_curious]

@arunnalpet

not clear on how each grid can detect an object which is larger than the grid

Applying YOLO anchor boxes

Applying YOLO Anchor Boxes A previous thread provides an introduction to extracting from labelled ground truth data a small set, on the order of 3 <= k <= 10 , of shapes for what is referred to in the YOLO 9000 paper as dimension clusters or priors, but in this class as anchor boxes . NOTE: Just to make things more confusing, the YOLO 9000 paper uses anchor boxes to refer instead to the hand-picked shapes described in the Faster R-CNN paper, not the ones generated by K-Means. The YOLO 9000 pap…

Look specifically for the paragraph containing…

(this works even if this results in values larger than the grid cell dimension – it is exactly the mechanism that allows YOLO to predict bounding box shapes larger than one grid cell).