8.0 Image classification

     In 'Computer Vision Materials' category, I am going to summarize the data about Computer Vision.

 

 

 

● Classification

     Classification is used in many ML parts, but today we are going to concentrate on image classification.

 

street light or not?

The object in the yellow bow a street light or not?

 

polata palace or not?

Not only general object, we can also find specific object. Like finding out whether the object is Potala palace or not?

 

type of scene

Not only finding out what objects in the image, we can find out what type of scene is it. In this case, we can come up with words such as outdoor, marketplace, city. (Type of Scene)

 

detection

We can detect one specific class in the photo of many objects. (Detection)

 

segmentation

Image segmentation, which breaks the image into pixel units and finds out which object class it is, is also possible.

 

 

● Challenges

However, image classification has several challenges to overcome.

 

viewpoint variation

Same objects look like different in various viewpoints. (Viewpoint Variation)

 

illumination variation

Same objects look like different when illumination changes. (Illumination Variation)

 

Scale Variation

Same objects look like different in various scales. (Scale Variation)

 

background clutter

Background clutter is a problem that background and object are mixed up so that it's hard to separate two.

 

Deformation

When the object can move or change its form, it's more difficult to figure out which object it is. (Deformation)

 

Occlusion

If the object is partially hidden, it is called 'Occlusion'.

 

Intra-class variation

'Dog' class has several species which have different size, hair, etc. Objects in the same class can look different.

(Intra-class variation)

 

 

This is a classical classification pipeline.

classical classification pipeline

 

In next lecture, we are going to learn about 'Bag of Visual Words'.