YOLO

https://arxiv.org/abs/1506.02640

For object detection.

Objection direction as a regression problem to spatially separated bounding boxes and associated class probabilities using a single neural network.

Advantages:

1. Extremely fast - since it is an regression problem.

2. Reasons globally about the image when making predictions.

3. Learns generalizable representations of objects.

Method/Process.

First, divide input image into an $S \times S$ grid. If center of an object falls into cell - cell is responsible for detecting the object. Each ggrid cell predicts $B$ bounding boxes and confident score for the boxes. - IOU between the predicted box and the truth (score)

Each box consist of 5 predictions. $x, y, w, h$ and confidence.

Each grid cell also predicts class probabilities $Pr(Class_i | Object)$.

At test time, we multiply conditional class probabilities with box confidnece prediction, the result is then

$$Pr(Class_i) * IOU_{pred}^{truth}$$

Training: details in paper. Final layer predict both class probabilities and bounding box. Final layer is linear activation, other is ReLu. Also dropout

Loss function: