Take Input image

Resize any input image to 448×448 pixels (This varies between models, its 640×640 pixels for a YOLO v8 for reference)

Then we split the image into S-by-S grids, the S in the case of YOLOv1 is 7×7 grids. So, each purple box above, will have 64 pixels inside it.

It’s going to get a little complicated but understand that each cell (purple box) is responsible for predicting one object!

In the image above, the blue cell is responsible for predicting the person and the red cell is responsible for predicting the horse, the cells represent the centers of each object. In the case of object detection, it’s the center of the bounding box if you look closely.

Let’s assume that position of the center for the person in this image is (200,311,142,250) which represent the coordinates (x1, y1, x2 or w, y2 or h). It’s pretty noticeable that we have some quite large numbers.

When looking at machine learning, having large numbers makes the problem much more complicated so we can attempt to normalize this position and make it “relative” to the grid that it is inside of.

We take the top left point of the “green” box in this case and transform the position of our center using these equations.

I understand that it might be getting a little complicated here, so let’s look at an example calculation.

After performing that, our center point transforms from (200, 311, 142, 250) [Relative to 448×448 image] → (0.13, 0.87, 0.31, 0.56) [Relative to its grid cell] which makes it much easier for the machine learning model.

In terms of a training perspective, these new values now will be the “targets” for the YOLO model to achieve. This single step is done for all 49 grid cells for any input image. This is also popularly known as label encoding.

In this case, we have two centers in two different grid cells, one for the horse and one for the person, these will show up as the values marked in green in the pic above. For grid cells with no center (object) inside then, the output will be 0. This process is done for all 49 cells.

We can understand the first 5 values in the table. Additionally, p1 to p20 represent all the classes present in the pascal dataset preprocessed by using one-hot encoding.
Example: If my classes were (Elephant, Person, Horse, TV, Computer) then the grid-cell which has the center of the person class would have the vector (0.0, 1.0, 0.0, 0.0, 0.0).

At this point, pat yourself in the back if you’re able to understand all this (took me a while).

After performing all of that computation, we have the transformed an input image into comprehendible training data for the YOLO model.

You might have been wondering what the 5th column in the table represented (I did not clearly mention it anywhere). This represents the confidence of the model at its prediction, ranging from a 0 → 1.

Let’s understand the output prediction vector so this makes a bit more sense.

The meaning of a “Prediction Vector” is essentially the output we get out of the model. It is important to understand that one grid cell will output two bounding boxes (importance of this will be explained shortly). Totally, with everything, a single prediction from a model will have 30 parameters, which will look something like this.

First 5 values belong to the first bounding box, second 5 belong to the second bounding box, and the last 20 belong to all the class probabilities.

Now we need to understand how the output is “parsed” to get the final bounding box out of two predicted bounding boxes.

Each bounding box has its associated confidence (the 5th value). This is multiplied by the maximum value of the p vector to give us the class confidence. Class confidence (C_hat) = confidence * max(p_vector).

The bounding box with the higher-class confidence is chosen.

Take a breather and understand what we have done until now

YOLO v1 has a surprisingly simple architecture that was inspired by Google Net.

BBox Loss

Object-ness loss

Classification Loss