Is AI labor intensive?

Contrary to popular perception that AI will cause disruption of labor and employment, it is extremely labor intensive.

It is not blasphemous.

Ask those who have worked on Computer Vision AI projects, they will vouch for ‘data label’ being the MOST important factor for success. 

Most computer vision AI projects range,

from object detection 

to facial detection

to tumor diagnosis

to self driving cars

Accurate & efficient labelling/"annotation"for each type of business problem is very critical for the success.

As per McKinsey's report towards the end of last year, amongst the various AI functions, robotic process automation, computer vision, and machine learning are most commonly deployed. 

Types of Annotations

The annotations vary based on the type of use case being addressed.

a) Classification

This is the most commonly used type where humans label images manually by marking whether an image consist of car / tree / certain specific objects for which the use case is being developed. Tech giants like Facebook and Google use this regularly with its own users to label captcha images.

b) Outline annotation

This type of annotation is where one tries to identify x, y coordinates by drawing shapes around the outline of the object of interest. There are four approaches to outline the object

• Bounding Box

In this approach 2D boxes are drawn either manually or by tools and then labeled for training for Deep Learning use cases.

This approach is quite cost effective and we can label multiple objects in an image. But it cannot label an image with a meandering river for example.

• Polygon

Polygon labelling approach is a more exacting approach but is more time consuming for labelling, as shown by figure-eight in one of their blogs. This type of annotation is also used for labelling medical image scans e.g. lung cancer nodules in MRI scan.

• Dot

This type of annotation is used for counting jobs and gesture or facial recognition tasks. For example, for counting the number of people in stadium through drone footages, the labeller needs to put dots on each individual in the image.

For gesture and facial recognition, the dots are placed on the face to ascertain the emotion, interest level of person for a prop / product in retail store etc.

c) Pixel Labeling

This technique is used to label each and every pixel in an image. In this technique, an image is divided into multiple segments. Once segmented, polygons are drawn around objects of interest. But a very detail level of information is labeled including coded RGB pixels. Due to the high level of precision this technique is supposedly most expensive and time taking

How the world is addressing?

A plethora of solutions and initiatives have been undertaken to address the annotation challenge.

ImageNet has around 15 million images exactly for this purpose and is quite popular for initiating in Vision AI. But, it suffers from large label ‘boxes’, few label types and high error rates. 

Large companies use ‘subtle’ crowdsourcing, either using captcha image labeling (Facebook) or using Human Labeling (Google). Some startups like Figure-Eight, Handl, LabelBox use a hybrid approach for labelling.

In China, call center executives are queueing up to do manual labelling,

A full-time data-tagger at BasicFinder can earn 6,000 to 7,000 yuan a month, along with accommodation and social benefits. In the first three quarters of 2018, the disposable income per capita in Beijing was 46,426 yuan, around 5,158 yuan a month, according to local government statistics. (Source: xinhuanet)

On the API front, Amazon Rekognition provides Storage and Non-storage APIs for images and videos. Google Vision API provides ability across six categories like label detection, text detection, face detection, landmark detection, logo detection and safe search detection.