[Security News = Dong-Hyeon Lim, Supervisor of Smart City Service Team, Seocho-gu Office] In August 2020, Seocho-gu, the Ministry of Science and ICT and the National Information Society Agency, joined CPRO in the ‘Artificial Intelligence Learning Dataset Construction Project (Secondary)’ free task section. NamuPlanet formed a consortium with the National Safety Competency Association and was selected as the highest score, and the data set is being built until February 2021. 

▲ A view of Seocho-gu Office [Photo = Seocho-gu Office]

The theme proposed by Seocho-gu is 'building a dataset for measuring congestion in urban areas of vehicles and people using BirdEye-View', a characteristic of downtown. Seocho-gu leases the rooftops of 20 landmark buildings in the building, collects high-performance wireless CCTV images (Bird Eye View) that are installed and operated, and plans to process them together with a consortium to create a reference dataset for improving AI engine performance such as counting and congestion and density.

▲ Screen shot of the result for the AI learning dataset construction project (2nd) [Photo = Seocho-gu Office]

The government has defined a job group called labelers as the “Data Dam-New Deal Project,” including the AI learning dataset construction project, and promised active support. This project is the first in the case of a data set construction project in which a local government has become a supervisory organization, and I would like to summarize the contents of the Seocho-gu Bird's Eye View data set as follows in order to establish future government policies and develop the industry.

Start of construction business, “data acquisition”
Seocho-gu acquired Bird's Eye View by installing 40 fixed cameras on the roof of a landmark building to secure the bird's eye view angle, and made it the starting point of the business. As shown in Pictures 2 through 5, the main targets are places where actual congestion occurs, such as Sadang Station, Gangnam Station, Express Terminal, Gangnam-daero, and Gyeongbu Expressway, as the main goal, and designed to be a meaningful dataset in terms of congestion. To reflect the various needs of the industry, the image format (photo and video), image size (2M pixels and 5M pixels), compression method (H.264 and H.265), distance (far and near), shooting angle (based on ground level) From 30 degrees to 90 degrees), object size (from the size where personal information cannot be checked to smaller), etc., various angles of view were set and high quality images were acquired.




▲ Bird's Eye View (Picture 2~5 clockwise from the top left) [Picture = Seocho-gu Office]

'Data processing' for personal information protection
The acquired data may include personally identifiable information. Therefore, unlike the existing file distribution work method or file download work method, the basic environment was implemented as VDI (Desktop Virtualization: Virtual Desktop Infrastructure) to present a fundamental solution to personal information distribution and personal information infringement. In other words, the labeler complied with the law by allowing all administrators to monitor and evaluate quality only in the VDI environment, using the VMware Horizon VDI client to access, work, and store (not copy) the Seocho-gu Birdeye View platform. As shown in Figures 6 to 11, the working platform in the virtualized environment is developed based on open source CVAT, so that the labeler works intuitively, adds convenience to inspectors over three rounds, and is designed to secure deep visibility to managers. For specific images and images that identify vehicle numbers, such as photo 12 to photo 15, only the relevant area is masked to minimize the damage rate of original and learning data.


▲ Clockwise from the top left, picture 6 after VDI execution, picture 7 platform screen, picture 8 worker screen, picture 9~11 manager monitoring screen [Photo = Seocho-gu Office]



▲ From left, photo 12, 13 Annotator work result (person) [Photo = Seocho-gu Office]


▲From left, photos 14 and 15 annotator work results (vehicle) [Photo = Seocho-gu Office]

‘Providing data’ that will satisfy user satisfaction
The Seocho-gu consortium is based on images and metadata (photo 18) labeled for each 300 hours for people and vehicles, 14,400 hours of bird's eye view original video, and video compression technology (H.264/H.265) for differential use to improve AI reliability. 1,200 background images (photo 19) will be provided as a result. Legal review is underway with actual videos to check whether personal information can be identified in all videos, and only the reviewed data is uploaded to aihub.nia.or.kr after inspection through TTA is completed. The object flow direction, environmental information such as fine dust data, and distance and angle data between the angle of view and the camera's distance and angle data measured by a laser meter are included in the metadata so that users' satisfaction with the use of the dataset can be improved. For reference, the Seocho-gu consortium is jointly planned by a HW manufacturer (Cpro) and an artificial intelligence service company (NamuPlanet) from the perspective of consumers, and is trying to become a 'sufficiently useful dataset' in the current situation.



▲Photo 16 Open Talk(Kakao Talk) asking an annotator question (left) and Photo 17 Open Talk standardizing annotator question results (right) [Photo = Seocho-gu Office]



▲ Photo 18 labeled meta data [Photo = Seocho-gu Office]

▲Photo 19 Background image for difference recognition [Photo = Seocho-gu Office]

'Data utilization' in various areas
The data set prepared in this way can be used not only as the reference data for the AI counting engine and the AI service for measuring congestion (dense), but also as reference data for various empirical services in the ITS area. In addition, since it was not labeled because we did not think about it at the moment, considering that new labeling demand may arise at any time, a sufficient amount (14,400 hours = 30 angles of view * 20 days * 24 hours) of high-quality original data is also available on aihub along with labeling and metadata. Will be provided through.

'Proof as a service' through pilot service from January
The Birdeye View demonstration service selects two birdeye view angles of view (small) in Seocho-gu and overlays the results of real-time artificial intelligence congestion analysis on the original stream, and from January to the web (birdeyeview.seocho.go.kr) and the app (Seocho Smart) City) will operate a pilot service. In particular, by demonstrating the corona response convergence service, the result of real-time artificial intelligence congestion analysis will be linked to media such as SIP broadcasting terminals and electronic signs to transmit periodic announcements according to the congestion level, thereby demonstrating the possibility of using a dataset that contributes to national health.

Suggestions for improving data set reliability and quality
Since the metadata required by AI service development companies may vary, it is necessary to collect a lot of various original data apart from labeling. This means that the data set construction project needs to be shifted to the perception that it is a business to purchase original data from a certain point of view, and all involved must agree on responsibility for the reliability and quality of the data set.

Where you plan your business, you should try to create a neutral dataset. Associations representing the industry (○○○○ Research Association, ○○○ City Association, ○○○○ Technology Association, etc.) should also participate in the planning to create a wider and more universal standard and broaden the scope of data. In particular, there should be no failure to participate in the data set construction project or drop out of the data set construction project because specialized companies cannot be grouped as in the secondary project.

The diversity of the members of the participating consortium is also required. The AI ecosystem that everyone wants when the dataset business should not be concentrated on a few specialized companies operating a labeler and emerging AI SW companies, but rather should be a service provider (hardware manufacturers to companies that apply AI services) so that the artificial intelligence ecosystem that everyone wants can be created.

I hope that public institutions will also be interested in and actively participate in dataset projects. In particular, if high-quality data is required, the public must participate in the data acquisition and purification process. If the host institution is a public institution, it will not self-help the decline in data reliability, so you can think of high-quality reflection benefits.

It also seems necessary to improve the morality of participating companies. In fact, I hope that results will not be submitted with learnable data acquired through inappropriate channels such as YouTube or Chinese metadata purchase, or will obscure the purpose of the project due to inaccurate labeling and hinder the development of the AI ecosystem.

For reference, inexpensive video datasets that are already in circulation in the market or that can be easily purchased in China, even if it is not a legal issue, result in 'only insignificant amounts', 'not unique', and 'statistical errors' from the perspective of artificial intelligence. It must be filtered.

Lastly, the job of labeler is a hopeful job that can be given to those who are marginalized from society and those who are difficult to re-enter from the perspective of public welfare. I hope that the dataset business will become more active, and above all, appropriate treatment should be defined and guaranteed. The Seocho-gu consortium pays about 100,000 won each based on an expected 8-hour workday and is striving to save the purpose of the business.
[Written by Dong-Hyeon Lim, Supervisor of Smart City Service Team, Seocho-gu Office]

source : www.boannews.com/html/detail.html?idx=93760&tab_type=1