Object Detection

Outline

- Sliding Windows

- Bounding Box

- Bounding Box Pipeline

- Score

Image classification predicts the class of an object in an image

Classification and object localization

--> Locate the presence of an object and indicate the location with a bounding box and their classes

Sliding Window

= algorithm

- If we want to detect a dog, we consider a fixed window size

- If chosen property, the dog will occupy most of the window

Essentially a sub image that we would like to classify as a dog

The other sub images - classified as background

(Image that does not contain the dog)

*Process

1. Start in one region in the image, classify that sub-image

2. Then shift the window and classify the next sub-image

3. Repeat the process -- when the object occupies with of the window, it will be classified

Problems of Sliding Windows

1. Overlapping Boxes: object detects often output many overlapping detections

2. Object Sizes: have the issue of object sizes, where the same object can come in different sizes/Solution: reshaping the image

3. Overlapping Objects: this may pose issues to the sliding windows

Bounding Box

Bounding box = a rectangular box that can be determined with the lower-right corner of the rectangle with coordinates y=0 and x=0

Y and X are not the same as the classification labels y and the image x

Upper--left corner = (Ymin, Xmin)

Lower--right corner = (Xmax, Ymax)

They are just to illustrate the coordinates of the Bounding Box

--> The goal of object detection is to predict these points, so we add a "hat" to indicate it's prediction

Bounding Box Pipeline

Like classification, we have the class y and x

- we have a dataset of classes and their bounding boxes

Similar to classification, we use the dataset to train the model; we include the box coordinates

--> result: object detector with updated learning parameters

Input the image with the objects we would like to detect

We have the predicted class and the box coordinates

Score

- Many object detection algorithms provide a score letting you know how confident the model prediction is

Each column in the table has an image and it's prediction

The first row: the score ranging from 0 to 1

The second row: the class

The third row: the image and its bounding box

For the first row, we see the prediction is dog

--> but the image does not look like a dog

As a result, the score is 0.99

= the model is confident about its prediction

For each detection, a score is provided, we can adjust so we only accept detections above a specific score

*Usually models will only output objects over a specific threshold*

Copyright Coursera All rights reserved