News
Some decisions directly impact the investments and functionalities of solutions that use Computer Vision. The camera model used to capture the images and the recurrence of the episodes to be observed are some of the points capable of significantly increasing the final value of the project, which must follow the real needs of the company. Image: Computer vision solution: do you know the costs?
After understanding the multiple possibilities that Computer Vision solutions provide, the next challenge for Pix University is to have managers dive into the cost drivers, understanding which financial resources will be needed to put the project into execution. Haven't you learned about Computer Vision yet? Understand it here . First of all, it must be made clear that the sky is the limit when it comes to solutions based on Artificial Intelligence (AI) . However, throughout the creation of a Computer Vision system, a series of decisions can be made, each of which can increase or decrease the final investment of the solution. We cannot forget that every project must be scalable, that is, it needs to be executed sequentially so that it brings savings in financial resources over time. This means that, even if the initial investment is high, the solution will bring a significant reduction in the costs incurred by the company, promoting advantages in different aspects. As superior as they may seem, certain choices will not always be essential for the final result to be achieved. We will now take a journey through the main points that impact the financial value of solutions, keeping in mind that the focus must be on the real needs of the project.
Speak with one of our specialists and discover how Pix Force can transform your business
Defining the database that will power the system
Solutions involving Machine Learning (ML) are designed with the help of algorithms capable of learning new patterns and repeating them. However, for the algorithms to work, a database is needed to power them. And it is precisely these data sets that we call datasets. Understand what Machine Learning is and how it works here. The more customized the project, the greater the investment required to execute it. Solutions have their investment reduced if they use datasets that already exist in the market and are easy to acquire. However, there is a challenge related to the adjustments necessary for the dataset to work according to the needs of the solution being developed. This happens because ready and available datasets tend to be generic, after all, they were not produced for that specific function. This is the case with people walking on the street, which can be easily found through Google search engines. In this case, two things can happen: either the application truly needs generic data, such as crowd detection; or to use such datasets, it is necessary to spend a bit more time adapting the data to the project, making adjustments related to specificities such as the color and lighting of the images. For the first case, the dataset cost becomes extremely low; for the second, a bit higher than the previous one, but still one of the cheapest ways. When the company does not have the dataset but can easily produce it, we have an increase in the process cost, although it is not as significant. This is the case for solutions where it is possible to use cell phones, security cameras, or low-cost models to capture the images that will be used. For example, images of bottles passing on production lines can be easily acquired, without the need for large investments for capturing. The process becomes more expensive when it is necessary to build the dataset, as is the case with laboratory environments created by Pix Force for inspecting spoons. There are also kits aimed at data collection that can be deployed in a single day, which makes execution easier and cheaper. However, if it is necessary to acquire a camera capable of matching the data processing speed, the investment rises exponentially. Lastly, the most complex and consequently most costly situation occurs when the event to be observed is in a dangerous location and/or has low occurrence. In these cases, it is necessary to create a data collection plan which, on its own, can become a specific project.
Image capture: the impact of cameras on project investments
Another fundamental point in the project's architecture is related to the choice of devices that will be used to capture the images. Of course, each scenario may demand the use of certain types of cameras; however, there is a basic premise that must be observed to avoid waste: the more sophisticated the device, the higher the cost of the solution. Understand here how image capture is done in Computer Vision. There are projects that can be carried out with the help of mobile cameras, which are easily accessible and low cost. Obtaining images through cell phones, however, can only be done when lighting and capture angles are not relevant to the final result. Another possibility is the use of security cameras that are already installed, and their images can be used for the formation of the project. It works well for counting, but not for measuring object sizes with precision. In both cases, the capture devices represent a significant financial saving, but they can only be used when obtaining the data does not demand high levels of sophistication. In the case of cell phones, each person can capture the image differently, with large variations in lighting, angle, and framing. There is a saving in hardware, but the costs of processing and programming can increase depending on the quality of the data to be obtained. There are market cameras superior to security ones, capable of delivering interesting results while keeping costs reduced. This type of device usually offers the best cost-benefit ratio if used in a controlled environment, that is, without large variations. On the other hand, advanced models capable of working at very high speeds increase the project's cost, and can be used in very specific cases.
The higher the recurrence, the lower the price
Thinking about the recurrence of episodes that need to be observed also impacts the costs of the solution. A basic example can be found in projects that need to capture images of stains on fabrics: the higher the recurrence of stains, the easier it will be to create a neural network capable of detecting the episodes. Still don't know what they are and how Neural Networks work? See here. Imagine the following situation: the system has to be able to detect and classify stains on a fabric production line, whether they are oil or ink stains. If we have an average of 100 stains per day, with half of each type, in a short time we will have enough data to create a neural network that knows what the stain is and can classify which type it is. However, if we have 1 stain per day, it will take us months to have enough data to classify the stains, so the approach changes to anomaly detection. In other words, we train the network to say that anything different from the correct one is wrong and we detect that. The issue is that we won't be able to tell what type of stain it is until we have enough dataset, but the defect is found. From an investment point of view, anomaly detection requires less effort than classification, however, it must give up one piece of data, which is the classification of the anomaly. It is known that there is a stain, but not what type of stain. When there are not enough samples, the dataset collection process can extend for months, increasing project costs. This is the case for solutions aimed at identifying leaks in gas pipelines, where errors will not be triggered just so that anomalies can be identified. The frequency of episodes could, therefore, define the approach used in the project.
Processing the data: time required for image analysis
The next step that needs to be considered is related to image processing. It is possible to analyze the data in real-time; however, the faster the response, the more expensive the process will be. The response time can vary by seconds, minutes, or hours, and the choice must be made based on the actual needs of the project. It is necessary to evaluate whether the image analysis really needs to be done online, with immediate results, or if there is the possibility of working with offline systems capable of considerably reducing costs. Another point that must be analyzed is whether the data can be processed in the cloud, provided a good internet connection is available. The initial cost of the cloud is much lower compared to hardware investment, but it is necessary to take into account the maintenance value that will need to be spent on a recurring basis. There are Artificial Intelligence (AI) projects where image processing servers require investments that reach around R$ 100,000. In contrast, the cloud maintenance cost can reach R$ 2,000 or R$ 3,000, which leads us to rethink the pros and cons of adopting cloud processing. Another relevant factor of the cloud is related to data flow: more expensive than processing 1 gigabyte is the action of sending and receiving it back, as the flow is more costly than the processing itself. Each of these decisions must be evaluated carefully, as changes in the trajectory of projects can become unfeasible. Remember that the goal must always be to meet the demands of the process. We all like immediate results, but is that really necessary for the project? Investments for responses in 1 second or 1 minute can define the financial viability of the project.
Information storage
Storage on internal servers is cheaper than in the cloud, but there are some issues that need to be considered. First, if the entire processing structure was built to happen in the cloud, data storage can also be done in the cloud. Another issue is the storage period, which becomes more expensive as it extends. Cloud storage can be done on SSD or HDD structures. In the first case, data is accessed extremely quickly, but at a much higher cost. In the second case, the speed is slightly lower and the price drops considerably. If Pix Force's cloud is used, there will be a recurring cost for data maintenance. However, if the company needing the solution decides to use its own cloud, there will be costs linked to data integration, after all, it is complex to allow external people to access your cloud. There is also the possibility of local storage using the company's computers, a cheaper option that may be viable.
Two possible scenarios: the most expensive and the cheapest
To make everything even clearer, let's think of two scenarios: the most expensive project possible, which we will call A, and the one with the lowest cost, which we will call B. PROJECT A: Stain detection on fabrics. Process at 20m/s. (Requires high-speed camera), processing response in 1 second, aggressive environment, therefore processing must be in the cloud. 1 failure per day on average, the stain must be classified as grease, ink, or fabric weave failure. PROJECT B: Counting cement bags. Response 1x per day, use of cell phone, processing inside the device itself, constant data availability, storage on the smartphone itself. We are talking about something around R$ 600.00 per month for project B and R$ 200 thousand plus R$ 5 thousand monthly for project A. These are illustrative values only. In scenario A, there is a need for an image capture system with a superior camera, with advanced speed and high market value. The recurrence rate of the episodes that need to be observed is very low, which will require a lot of time for the database collection to be completed. Still in project A, we need an online response, requiring a high-power computer dedicated to real-time processing. Storage is done within the company's own cloud. In scenario B, considered the cheapest possible, the company already has the images, captured with its own camera. Recurrence is high, meaning that at all times we have episodes of what is desired to be observed. Access to the images is easy, as is handling the camera, without the need for external teams to perform the task. Still in this scenario, we have a response time of five seconds, offline, with no recurring cost, image processing performed internally (locally) with the smartphone itself. In addition, there is no need for data storage, and whatever needs to be stored can be done internally. We realize that there are a series of variations that need to be taken into account when structuring a project. Some points will have a considerable impact on the final outcome of the solution and, therefore, open up space to save on physical and financial resources. Regardless of the choices, the fact is that new systems can be created to revolutionize emerging pain points, with Artificial Intelligence (AI) being an inexhaustible source for increasing the competitiveness of different market segments. Keep learning: understand the future of Computer Vision here.
Fabio Caraça
Fábio Caraça is the Chief Growth Officer at Pix Force. He leads Pix Force's transformation into a scalable SaaS operation, combining strategic vision, culture, and high-impact execution.
