With the development of computer hardware technology and the explosion of data, but also gave birth to the application of AI technology penetration in all walks of life. And want to apply AI techniques to all walks of life, the data is a necessity. Because the data model directly affects the quality of AI eventually trained. AI modeling threshold is not much, but the data is the real threshold. Therefore, the industry circulating passage number of artificial intelligence, there are that many artificial . The previous describes how data tagging and attention today to learn about what types of data tagging?
Who will do the tagging data
At present each company need to mark data is generally done with the next few large companies or persons
The company own or trainee recruiters to mark
The biggest advantage of this approach is: You can check the quality and labeling of progress at any time for quick communication and adjustment, the data can be done confidentiality is not compromised; disadvantage is the higher cost
Outsourced to data labeling company
The biggest advantage of this approach is: very fast, low cost, companies also have a certain label markup tools development capabilities, strong customization capabilities; but the disadvantages are also clearly marked the company was mixed, uneven, you need to always check the label quality, greater rework costs. In addition confidential data can not do is not compromised
Outsourced to private parties
Advantages and disadvantages of this approach is basically the same label company. Advantage is lower cost, because you can recruit a large number of people, such as rural idle staff, part-time personnel. Training can be put into a slightly lower mark.
Who manages the data labeling
For different data assigned to the person, the appropriate role can be divided into three types:
Data tagging staff
Responsible for labeling and summary data
Data inspectors
Is mainly responsible for the data denoted by QA, often used way to check sampling, sampling may be subdivided into portions (a ratio of about 20% to 30%) or all of the sampling
Data Management
Responsible for personnel assignments, schedule follow-up, labeling and external training / internal coordination and communication, etc.
Type of annotation data
AI technology due to the particular example scenarios will be very different, so there are many types of label, as shown in detail as follows:
Image annotation -2D marked border
This should be the most common type of marked way, used to detect a target object corresponding area marked block at the target object around, as shown below:
-3D image annotation label border
Also known as the cube mark, compared to 2D label, it can also show the approximate depth of the target object. As follows:
Image annotation - Semantic Segmentation
Depending on the detection region of the image pixel labeled as different, as follows:
Image annotation - Polygon label
According to the shape of the demand labeling of the target object, often used to not use irregular target object frame mark, the need for tracing point in each of the key points of the target object, regardless of the ultimate what shape are to be able to react outline shape of the target object and all edges, As follows:
Image annotation - straight or curved labels
The labeling of the target object corresponding to the demand position line, the line may be straight or may be curved, commonly used in the partition boundaries indicate things. Commonly used in the autopilot, as it follows:
Image annotation - point mark
This label is generally used for face recognition, body posture tracking (such as POSE algorithm), etc.
Video Labels - tracking label
Tracking target object in a video labels or continuous image forming trajectories associated with ID
Text labels - in Chinese and English voice transcription and proofreading
Chinese English voice-to-text English or Chinese text-to-speech.
Voice annotation - customer service voice annotation
Outbound robot outbound call record a voice annotation success or fail, then surgery training.
Common annotation tool
Highlighter used as follows:
labelImg
Download: https://github.com/tzutalin/labelImg
Note: When using labelImg, the path can not contain ChineseElf marked Assistant
Download: http://www.jinglingbiaozhu.com/