Shulou(Shulou.com)06/03 Report--

The author explains his own work experience by analyzing the requirements of ToB image AI products.

The author hopes to discuss with you how to complete the demand analysis of ToB image AI products. The content of this paper only expresses the author's habit of product demand analysis in practical work. Welcome to communicate.

Product requirements analysis is a required course for product managers. Only by thoroughly understanding the requirements can we avoid errors in later product design. For ToB AI image products, the author generally follows the following analysis routines.

First, business requirements analysis

If you do not look at market analysis, then business requirements analysis should be the first step for product managers to enter product planning.

In particular, ToB products, only find out the business needs of users, the products can be more suitable for users, so that users can not stop to pay the bill, the author of this analysis is generally from the following core points to collect user data.

1. Business requirements background

First of all, the author is generally very concerned about the need to figure out the business background of the current requirements.

Collecting business requirements background is not only to find out why users do products, that is, what bottlenecks in the current business state need to be solved by AI technology; secondly, to clarify the business requirements background is also conducive to the collection of industry attributes, so as to prepare for later competitive product analysis.

two。 Business scenario

As we all know, although computer vision has made great progress, there are still many bottlenecks, such as camera angle, lighting, shadow, occlusion and so on. Therefore, in the demand research, it is necessary to sort out what scenarios users want to use image recognition technology.

Taking the logistics industry as an example, the author does not consider the hardware such as camera, bandwidth, server and so on, but only considers the scenario and the algorithm application of the corresponding scenario:

Security scene: detect the human body under the warehouse under the security scene and identify the human body attributes. The difficulty is that the height of the security camera set up in the logistics warehouse is much higher than that of other scenes, and the scene covers a wide range, even if occlusion is not considered, the target will be relatively small. At this time, the detection difficulty will be relatively increased, and there will even be the problem of small target detection. If you still need to identify human attributes (such as clothing, etc.) at this time, it will be more difficult.

Fixed camera scene: sometimes the overall solution needs to be considered according to business requirements, and an important part is the camera erection scheme. At least the camera erection angle, vertical distance and horizontal distance need to be considered in order to get a clear picture, which is beneficial to algorithm detection and recognition.

Special business scene: mobile phone shooting and security check machine The reason why mobile phone shooting is special is that because the photos taken by mobile phones are different because different people use different mobile phone models, the resolution of the pictures taken is different, and the angle at which the target is to be identified is also different. the time-consuming and concurrency problems caused by inconsistent resolution can also be solved by adding servers, interfaces and even data compression. However, the inconsistency of the angle will affect the accuracy of the algorithm itself, either higher false detection or more missed detection. In particular, the security check machine has different imaging results, so it will be difficult to study the algorithm. the image of the same target under the security check machine may only have two features: shape and color, which will increase if you want to complete the recognition. the image of the target under the security check machine is shown below (the picture comes from the network and is used only for description in the article).

Summary:

AI technology and scenario should complement each other. High-mature AI technology should be business scenario-finding technology, while low-mature AI technology should be more technology-seeking scenario. Why?

Because the scenario needs to be constrained in order to meet the feasibility of the technology.

3. Business requirements

After understanding the business background and business scenarios, the further core is to understand the business requirements. The simple understanding is what targets the users want AI to detect or identify, which is related to whether the algorithm will use video recognition or picture recognition in the algorithm requirements.

AI can not cover all the targets to be identified and blindly expand the scope of detection and recognition, which brings about at least two aspects: one is the acquisition of training data, and the other is that it is difficult to guarantee the final recognition effect.

All these will restrict the algorithm, which is also the biggest difficulty of current commodity identification (including the problem of similar sku). Therefore, combing the business requirements is conducive to the normal development and follow-up promotion of the algorithm. Take the OCR ticket identification as an example to illustrate how to dock the business requirements.

The requirement put forward by the business side is generally "We need to identify the so-and-so ticket", and the AI product needs to analyze the business requirement and convert it to the AI algorithm requirement as far as possible:

Determine the type of ticket to be identified, train ticket-> (red? Blue), invoice-> (printed invoice? Handwritten invoice? ) to determine the content to be identified, including two aspects: the first is the content of the ticket itself, such as numbers and different languages; the second is to clarify the field information to be identified. such as train tickets need to identify the name, date, price, seat type, departure station and final arrival station and so on. 4. Business value

After a lot of analysis above, I finally figured out where to do it and what needs to be done. Then, the next step is to analyze whether it is worth doing. Business value analysis is an important condition to measure whether the product is worth investing, and it is also the evaluation condition for the follow-up pricing strategy. The most important thing is to report the needs to the boss.

It can be analyzed directly and indirectly from the AI product side, and the direct analysis is the amount of usage generated by users every day after the service is provided. after all, the most commonly used pricing strategy is evaluated according to the amount of call; indirect analysis is the business value generated by users calling AI service, which can save labor costs, improve audit efficiency, improve click-through rate, and so on.

Of course, the best way is to quantify the analysis, the numbers can give the boss intuitive persuasion.

5. Business proc

After all the above contents have been analyzed clearly, we can basically assess whether this can be carried out as an AI product before reporting to the boss. If it is determined that it is going to be carried out, it is necessary to sort out the business process.

First of all, docking with users to sort out the current system process of the business, you can focus on which step the original system uploads the picture and how it is reviewed after uploading the picture. With these two pieces of information, we can basically make it clear when the AI identification service should be called, and how to optimize the next steps after the original system obtains the AI identification information.

The specific business process is not given here. The case can refer to the business logic described in "using AI technology to achieve online and offline interactive drainage".

Second, algorithm requirement analysis

Usually when the author communicates the requirements with the algorithm, the algorithm often asks the following points:

In what scenario; what needs to be identified; what is the cost accepted by the user; what is the performance requirement of the product (accuracy and time-consuming); the cloud is still local.

Combined with the above five points, the AI product manager needs to deduce the algorithm requirements:

1. The algorithm asks, "in what scenario?"

The product needs to describe the business scenario clearly. If it is based on the existing hardware solution of the business side, the image imaging conditions and image resolution should be included in the algorithm requirements. The best way is to get the pictures under the current conditions to the algorithm, one picture is worth ten thousand words.

In this step, the algorithm evaluates whether the current hardware conditions can meet the recognition requirements; if not, the AI product manager needs to do the following two steps:

(discuss with the algorithm, let the algorithm give hardware deployment parameters, including camera installation angle, camera installation height, camera model requirements, and other requirements (such as supplementary lighting, etc.); discuss with the business side that if you need to redeploy the hardware to implement the AI solution, whether the business side accepts it or not, the business side may need to evaluate the construction and hardware costs. two。 The algorithm asks, "what do you need to identify?"

In this step, take OCR as an example, you can refer to the above business requirements description. Under expansion, the algorithm requirement analysis not only tells the algorithm what needs to be identified, but also tells the algorithm what needs to be returned, such as the coordinate information of the returned target box, the confidence of the target, the number of targets, and so on.

3. The algorithm asks: "what is the cost accepted by the user?"

This generally occurs under the premise of the need to redeploy the hardware scheme, because the algorithm evaluation of hardware parameters is generally considered comprehensively, if the requirement is low-cost, the algorithm generally needs to sacrifice some performance to adopt a lightweight model and so on.

4. The algorithm asks, "what are the performance requirements of the product?"

Usually, the algorithm expects to achieve the accuracy and time-consuming that the user wants, and what AI products need to do is to confirm this value with the user. In general, some users will first do full research, and then will tell the product how much he expects.

However, many traditional enterprises access AI for the first time, and it is usually not clear how much this target can be achieved. AI products and algorithms and users first agree on an index to facilitate the work pilot, and then according to user feedback can collect error samples to quickly iterate the product.

5. The algorithm asks: "Cloud or local?"

This is related to the model scheme and hardware cost of the algorithm, as well as whether the product design is an API interface or a SDK service.

For example: according to the content to be identified in the business requirement description, the algorithm needs to evaluate whether to use video or picture for analysis. If video analysis is also needed, the network bandwidth cost will be very high. At this time, the algorithm needs to evaluate whether image analysis can be used, what is the lowest frame rate, what level of model to use, and the architecture needs to evaluate the network deployment scheme, whether the video server can be added locally for preprocessing, how much time is appropriate for transmission, and so on.

The product plays a project-driven role in this process, coordinating internal technology and users.

Summary:

The above content combined with business requirements analysis can generally put forward the requirements for the algorithm, and finally make it clear with the algorithm when to output the algorithm scheme.

PS: algorithm requirement analysis requires AI products to constantly communicate and confirm all kinds of information between algorithms and users (sometimes very dragging, because users can't predict their expectations before accessing AI).

Product demand analysis

AI products need to break the limitations of traditional GUI. The product form provided by AI products can include not only the complete system of front and back end GUI, but also the form of API interface and SDK.

1. API, SDK class

In the previous article "how to make a SDK product", the SDK product is briefly described in the form of API. The product requirement document needs to include interface input, output, algorithm accuracy, error detection rate, missed detection rate, interface time-consuming performance and algorithm constraint rules.

Take image recognition as an example, the core fields that a product needs to define include:

Input: picture format-jpg, jpeg, png, etc.; Image transmission format-base64 or url;ROI area-quantity. Default can be the whole picture, the maximum number supported; ROI drawing-rectangle (left vertex + length and width), polygon (all coordinate points); Identification type-if the API supports identifying multiple content, this field can be added to specify the content to be identified by the algorithm. Other field information such as authentication and timestamp can be defined by the developer; output: the core information is the same as the algorithm requirements, but needs to be implemented in the interface fields, such as the total number of targets, coordinate information of each target, confidence, special field information for other analysis targets Algorithm accuracy, error detection rate, missed detection rate, the author suggests that it is best to analyze business indicators, take target detection as an example, usually the algorithm is measured by mAP, it is evaluated from the target dimension, but the user usually measures from the dimension of the picture, if there are errors or omissions in a picture, the user may think that the image recognition is wrong; therefore, the product needs to be clearly defined. Time-consuming performance. The time-consuming performance here refers to the overall time consuming of the interface, that is, the time it takes for a user to input a picture to return a result, which needs to be evaluated by both the algorithm and the development. The product only needs to define the product requirements. 2. GUI class

If it is the form of the GUI class, the product should focus on how to design the product prototype, user experience, user use process, etc., which is no different from the traditional product design, just need to separate the algorithm requirements to the algorithm partner. At the same time, you can agree on the rules of the algorithm, and the so-called algorithm rules accept the constraint rules for users under the current algorithm capabilities, such as the format and naming of the files that the user is required to upload. What are the restrictions on the recognition results returned by the algorithm to the user?

As a product is actually very helpless, in principle should be user-centered mining users of the most natural user habits, but in the case of immature AI technology, the need to sacrifice some user experience.

Conclusion

The biggest difference between AI products and traditional products is that there is another layer of algorithm between products and development.

AI products can only identify the algorithm requirements first, and then further consider the product requirements after the verification results of the algorithm feasibility. Of course, the algorithm requirements are also part of the product requirements.

Http://www.woshipm.com/ai/3059898.html

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.