한국어 English Tiếng Việt bahasa Indonesia Unisem
한국어 English Tiếng Việt bahasa Indonesia Unisem
Subject

DATA LABELING

Name Admin
Date 21-04-01 13:31
View 228

Contents


 

DATA LABELING

Data Labeling
This month, as a Vision AI-based solution provider, we would like to introduce Data Labeling. Accurately labeled data sets are the raw material for the machine and deep learning revolution. Training the next generation of artificial intelligence (AI) requires vast amounts of data. Correctly labeled images train AI systems to accurately differentiate between vehicles and pedestrians or between hats and helmets.
Figure 1. Data Labeling for Traffic Objects

Data Labeling Types: Manual Data Labeling versus Automated Data Labeling
How do you create the accurate and scalable data set that the industry needs? To answer this question, you should first consider manual data labeling and automated data labeling. The difference in this approach to data labeling points towards the creation of smart data sets.
Manual data labeling refers to the process by which a human annotator identifies an object in an image or video frame. Annotators look at hundreds of thousands of images to construct comprehensive and high-quality AI training data. Specific labeling techniques are applied to the raw data, depending on the needs of the developer.
The answer to “manual labeling versus automated labeling” is finding the right data labeling process for your AI project. The right labeling tools and a well-trained and professionally managed annotation workforce can be a powerful combination for today's innovators.
The automated data labeling process has the potential to overcome some of the problems caused by the difficult annotation cycles. After training on a labeled data set, engineers can apply a machine learning model to an unlabeled data set. Then the model should be able to predict the appropriate label for the new data set. Automated data labeling algorithms can be improved through human input: after the AI has finished labeling the raw data, the human annotator reviews and confirms labels. Then, correctly labeled data can take its place in the training data set. If there is a mistake in the labeling, the annotator can proceed to fix it. This modified data can also be used to train labeling AI.
The answer to “manual labeling versus automated labeling” is finding the right data labeling process for your AI project. The right labeling tools and a well-trained and professionally managed annotation workforce can be a powerful combination for today's innovators.
Data Labeling at UNISEM
Until now, UNISEM IoT division has been using manual data labeling based on human annotators in the vision AI training. For accurate object recognition, many images must be prepared, requiring a lot of manpower and a lot of cost and time. The role of the human annotator cannot be overlooked but collaborating with people has a human factor that causes discord. In the process of work, one annotator might perform labeling on two hundred images a day, while another completes only one hundred images a day, moreover, there is often a difference in the quality of the specified labels which brings inequality among annotators and complicates the assessment and rewarding estimation.
Automated labeling AI can manage most of the easily identified labels. This brings the advantage of greatly accelerating the initial labeling step. Human annotators only need to make minor corrections or inspections of automatically labeled images, which improves productivity.
Figure 2. UNISEM IoT Division's In-House Automatic Data Labeling Platform Dashboard

Since data labeling is constantly required to improve existing solutions or develop new solutions based on image recognition, UNISEM IoT division started developing an in-house automated data labeling platform three months ago. At the current stage, our labeling AI system labels one hundred uploaded images within 4 seconds. Although the automated data labeling platform at the beginning needs human help, it will continue to train and improve the autonomous labeling function, so most of the Vision AI training data set preparation work will be done automatically.
UNISEM'S NEWS
2021.03.10

UNISEM PARTICIPATES IN AI-BASED COASTAL SAFETY PROJECT IN POHANG CITY

Read more ▷
UNISEM'S IOT PRODUCTS
  
Facebook
Twitter
LinkedIn

Subscribe to our Newsletter

I have read and accepted Privacy Policy.