【Object Detection】YOLO simple explaination Part2

In last time, I explained until Multi Detection by CNN.
Let's start from where we left off.

5. IOU

Multi detection I explained at last time also has a ploblem.
It may detect one object multiple times.
・Detection of one object multiple times.

Quote: What is YOLO algorithm?$^2$

To deal this, we use a indicator called IOU(intersection over union). It indicate area overlapped of multiple bounding box.
・IOU formula
IOU = intersect area / union area

Quote: What is YOLO algorithm?$^2$

According to IOU score, we remove bounding box that overlapped and has lower plobability one.
The determination of whether or not there are duplicates is made based on whether the IOU is above a certain value (0.6~0.7).

・Removing bounding boxes

Quote: What is YOLO algorithm?$^2$

This technic also called [non max suppression], the problem multiple detections is resolved by this.

6. Detection multiple object in a part

The basics fo YOLO have been covered up to Chapter 5.
In Chapter 6, we consider to detection multiple object in a part.
・Multiple object in a part

Quote: What is YOLO algorithm?$^2$

YOLO deal it with increasing output vector from 7 to 14.

The first 7 correspond to Dog, next 7 correspond to person(We teach using teacher data to be such).
When we want to increase the max number of object detected in a part, we have to increase output size like [7,14, 21,...], it must be decided when making model. Of course, the size of output of CNN increasing cause calucuation cost and time increasing.

Summary

This is th end of explanation about YOLO. It is CNN model with several technic for training efficiency.
・YOLO

YOLO has many version, and those have several difference. However, the priciple is same, so you can understand those with less time.

Thank you for reading.

Reference

(1) You Only Look Once: Unified, Real-Time Object Detection
(2) What is YOLO algorithm? | Deep Learning Tutorial 31 (Tensorflow, Keras & Python)