Big data industry innovation service media
——Focus on data·Changing business
In the field of computer vision, the more well-known ones are SenseTime, Megvii, Yuncong, and Yitu. They are not only more famous, but also have higher income levels and valuations. Judging from the latest financial reports, these four little dragons are all in trouble to varying degrees, with weak revenue growth and huge losses that cannot be reversed.
The dilemma encountered by the Four Tigers is, to some extent, that the computer vision track itself faces some commercial implementation challenges. Or perhaps the field of computer vision itself is undergoing structural changes.
We are paying attention to a not-so-well-known computer vision manufacturer - Gelin Shenzhen. Judging from its latest financial report, although its overall size is not as good as SenseTime and Megvii, its development trend is relatively good. Next, let's take a deeper look at the company and see what sets it apart.
Steady revenue growth and initial profitability
In recent years, Green Shen Tong's revenue has continued to grow, and the growth rate is not bad. In some innovative fields, sustained and stable revenue growth is always the most important. High revenue growth, high R&D investment, and high marketing expenses are the typical "three highs" characteristics of an enterprise's business scale expansion stage. Under this premise, moderate losses are tolerable. Of course, there are also serious problems when extremely exaggerated huge losses can even trigger a cash flow crisis.
In the first half of 2023, Green Shenzhen's revenue was 157 million yuan, a year-on-year increase of 34.35%, and its net profit was 1.938 million yuan, a year-on-year increase of 115.28%. For comparison, in the first half of 2023, SenseTime Technology's revenue was 1.433 billion yuan, a year-on-year increase of 1.26%, and a loss of 3.123 billion yuan; Yuncong Technology's revenue was 164 million yuan, a year-on-year decrease of 58.16%, and a loss of 315 million yuan.
Green Shen Tong’s income
Although Green Shen Tong's revenue is much smaller than that of SenseTime and Megvii, its current development trend is relatively good. Moreover, Gelingshentong has turned a profit. Both net profit and net sales profit have been significantly improved. If this trend can continue, Green Deep Eyes will be able to achieve stable profits, which is of great significance to computer vision companies. In addition, Green Shen Tong's sales gross profit margin has also steadily improved in the past two years, which is a good signal.
Green Shen Tong's profit situation
From the cost structure, we found an interesting phenomenon. While the overall business scale of Green Shen Tong increased, its sales and administrative expenses also decreased. In particular, administrative expenses fell by 18.79% year-on-year, indicating that its operating efficiency has been improved.
But on the other hand, generally speaking, sales expenses and administrative expenses will expand with the expansion of the company's business scale, and there is a certain limit to cost reduction and efficiency improvement. For example, next year, there will be less room for further compression of various expenses of Green Shenpu. limited. The core of the company's growth lies in the continued expansion of revenue scale.
Green Deep Pupil Cost Structure
With visual products as the core, four industry applications are laid out
So, what factors drive Green Shen Tong’s growth? To figure this out, we need to look deeper into its business structure.
In terms of technical product architecture, the bottom layer of Geling Shenpu is the "Shentong Brain", which covers the data platform (data collection, governance, standards) and training platform (model training, optimization selection, management). On this basis, it continuously optimizes its various aspects. technology. The core of these technologies is the corresponding algorithm model, including 3D stereo vision, robot perception and control, large-scale cross-border tracking, etc.
Further up, there are three standardized products of Green Shen Tong for customers, namely Zhiyuan intelligent front-end products, Lingxi data intelligence platform, and Shen Tong industry application platform. Based on the three standardized products, we can build products for customers in different industries. solution. At present, it is mainly used in four fields: finance, urban management, commercial retail, and rail transportation.
Green Deep Eye Technology Product System
Dismantling the revenue and cost structure of Green Shen Tong, its revenue is 157 million yuan, mainly from artificial intelligence products, followed by technical services. In the cost structure, operating costs (mainly parts procurement) are 59.73 million yuan, and gross profit is nearly 100 million yuan. Various operating expenses are nearly 100 million yuan, including research and development expenses of 77 million yuan, followed by sales expenses of 28.6 million yuan, and administrative expenses of 16.97 million yuan.
Breakdown of Green Shen Tong’s income
To further increase profit margins, Green Shen Tong needs to reduce the proportion of operating costs, which requires it to optimize its supply chain and improve its bargaining power. At the same time, it can reduce procurement costs by self-developing some core components; in addition, in terms of sales and administrative expenses, Expenses can be further compressed by reducing costs and increasing efficiency, but there is not much room for improvement. R&D expenses are an important part of Green Eye's costs, and it is necessary to improve R&D efficiency and increase output per unit of input.
Three key issues that determine the future development of Geling Shenpu
The above is the basic business situation of Green Shen Tong. Next, we explore three key questions in more depth.
1. Is it time for 3D vision to appear?
Among many computer vision technology providers, GreenTon stands out for its firm bet on 3D vision. Although 3D vision technology has not been widely used in the past few years, its potential is huge. The reason why 3D vision did not become mainstream in the early days is largely because it requires high data processing and computing requirements, and it also requires more advanced chips and hardware support, resulting in relatively high costs. At the same time, early 3D vision algorithm models were not mature enough.
However, compared with 2D vision, 3D vision has incomparable advantages. First of all, 3D vision can provide depth information for objects to help achieve accurate object positioning and identification. Secondly, it is able to capture the structure of objects in the real world, allowing for more accurate modeling and analysis of objects. In addition, 3D vision can also provide richer spatial information to help more complex scene understanding and navigation.
With the rapid advancement of chip technology and the continuous optimization of software algorithms, 3D vision technology has attracted more and more attention. Its application scenarios have also expanded from a single field to various industries. For example, in the field of autonomous driving, 3D vision can help cars better understand road conditions and predict the actions of pedestrians and other vehicles. In the medical field, 3D vision can provide doctors with more realistic diagrams of the human body structure and help them perform accurate surgeries. In addition, 3D vision is also used in many fields such as construction, entertainment, and retail.
To some extent, the field of computer vision is undergoing a technological and industrial change from 2D to 3D. With the advancement of technology and maturity of the market, 3D vision has gradually developed from a cutting-edge technology to a widely used solution. Under this circumstance, Green Eye's early accumulation in the field of 3D vision is expected to give it a competitive advantage in this market stage.
2. Layout large models, but insufficient investment
In Gelingshentong’s financial report, there are some sporadic descriptions involving large models. For example, its deep-pupil brain can support model training with billions of parameters; the multi-layer feature voting mechanism based on Transformer improves the proportion of correctly matched point clouds and improves point cloud registration accuracy. Currently, GreenTun is also developing large image pre-training models for use in scenarios such as video classification, image classification, few-sample detection models, and few-sample event tasks.
Judging from the current situation, large models are not the focus of Green Shen Tong's R&D, and its R&D team has a total of more than 300 people and its annual R&D investment is only more than 100 million yuan, which is not enough to support its cutting-edge technology exploration in the field of large models.
The author is in " Chinese CV Corps, Are They in Danger?" ! "As mentioned in the article, large models are the evolutionary form of deep learning. At present, the results of large models are mainly in the field of natural language, and the star product is ChatGPT. From a technical perspective, there should also be large models in the field of computer vision. Moreover, combining NLP and computer vision technology is an important direction for building multi-modal large models.
However, there are still many technical problems that need to be solved to achieve this goal. Transformer is specifically targeted at the field of natural language. Whether a large model with hundreds of billions of parameters in the field of computer vision can be built based on Transformer is still unknown. To build such a large computer vision model, it is likely that the underlying Transformer architecture will need to be modified or even a completely different infrastructure developed.
Of course, if a computer vision manufacturer can solve these problems, first launch a large CV model with hundreds of billions or trillions of parameters, and achieve a qualitative leap in technical performance and application effects, it will definitely not shock the industry. Less advanced than ChatGPT.
Judging from the current situation, there are also several domestic computer vision manufacturers launching large model products, but their ideas are basically at the stage of imitating and following ChatGPT, and launching some tens of billions or even billions of parameters based on Transformer. A large model with nothing new at all, it is destined to fail to make waves in the market.
3. Security is the largest application scenario of computer vision.
Finally, let’s discuss the future development prospects of the computer vision track from the perspective of application scenarios.
The author has always believed that security is the largest application market for computer vision. But to activate this market, there is a prerequisite, which is the intelligence of security cameras.
In today's security systems, the core function of the camera is to record video, and then transmit the video to the background for storage and analysis. Video data, especially the video data generated by high-definition cameras, often has a huge amount of data. If the video data has to be transmitted to the background for analysis, the pressure on data transmission and storage will be very high.
If we think about it differently, the front-end camera has strong video analysis and processing capabilities. Most of the video analysis needs can be met directly at the camera terminal. Only the needs involving big data analysis need to be processed by the background, which will be extremely difficult. Greatly reduce data transmission and processing pressure.
For example, if a city issues a wanted warrant for a certain criminal suspect, the data is transmitted to the backend, and then identification commands are issued to camera terminals throughout the city. As long as the person appears within the field of view of a certain camera, the camera itself can complete the identification, comparison, and verification functions, lock the suspect, and then immediately transmit the address and corresponding video data to the background. Multiple cameras can also automatically form a local area network to analyze the suspect's movement trajectory, predict his next action direction, and start cameras in the prediction area in advance. In this way, the efficiency of the entire system will be greatly improved, while the cost of data transmission and storage will be significantly reduced.
Let’s look at another scene. A car accident occurred on a city street. The traffic camera immediately recognized that it was a traffic accident through behavior and scene recognition algorithms, and then reported the information to the traffic management background and synchronized it to the public security system. The camera can still Automatically adjust the angle and focus to record the accident scene clearly and fix the scene information as key evidence for later accident handling. Compared with the current way of handling traffic accidents (people involved in the accident call the police and wait for the traffic police to handle it), this new way is more efficient, buying valuable time for traffic accident handling and first aid, and the division of accident responsibilities is more clear. Clear and avoid later legal disputes.
To realize the scenarios in the above two examples, the camera must be truly intelligent and capable of identification and data processing. On the one hand, the camera must have a built-in chip with strong computing performance, and on the other hand, it must have a powerful built-in Identification, data analysis algorithm model. At the same time, the cost of smart cameras must be reduced through technological innovation and large-scale mass production. Only when the cost is low enough can large-scale commercial use be achieved.
Obviously, whether it is hardware equipment or software algorithms, current computer vision technology products are difficult to meet the requirements of smart cameras. But precisely because it is difficult, it is more valuable. Just imagine, if any computer vision manufacturer can solve these technical challenges and launch low-cost, powerful smart cameras, it will reconstruct the entire smart security market, and its development potential is naturally unlimited.
To a certain extent, the value of an enterprise comes from the market problems it can solve and the market needs it satisfies. The more difficult the problem to be solved, the greater the market demand to be met, the greater the value the company can create, and the higher its competition barriers will be.
★ Follow the Data Yuan official account and reply "Green Shen Pu 2023 Semi-Annual Report" in the background to download the full version of Green Shen Pu's 2023 Semi-Annual Report.
Text: Yicai Yanyu / Data Yuan
