Ajai Chemmanam
Computer Vision Tasks
Date: 2022-04-07
Written by Ajai Chemmanam

Computer Vision

Computer vision is a subarea of AI and ML which deals with images and videos. The algorithms usually helps the machines to understand the contents of the visual media.

Computer vision Tasks

Computer vision tasks can be classified into different categories

Image Classification

It takes in an input image & classifies it into a class label along with some metrics.

Advantages

  • Most basic computer vision tasks.

  • Requires smaller networks when compared to other tasks.

Disadvantages

  • The location of the class is not specified.

  • Cannot be used to determine the number of objects in an image

  • Images with more than one class cannot be predicted unless we use a multi label classification approach.

Object Detection

It takes in an input image and outputs the labels & location of the objects in an image using bounding boxes. The localization enables it to identify multi class classification & deal with objects with multiple occurrences.

Advantages

  • Provides the position of the object within an images.

  • Can be used to determine the number of objects in an image.

Disadvantages
  • It does not give the shape of the object in an image.

Image Segmentation

In Image Segmentation, we mark the presence of an object by pixel-wise classification using masks for each object. This can help us in determining the shape of an object in the image

Advantages

  • Provides the shape of the objects by pixel level classifiation between object and background.

  • Can be used to determine the position and number of objects in an image.

Disadvantages

  • Requires carefully labelled annotation over the shape of the object

  • Requires much more complicated models.

Image segmentation can be further divided into

Instance Segmentation

The pixel wise classification is instance wise. Pixels from different instances of an object will be classifies seperately.

Semantic Segmentation

All pixels of the same class objects are segmented into on1.

Panoptic Segmentation

Both objects and background are segmented out in Panoptic Segmentation.