Play with the one-stop plan of cann target detection and recognition [basic]

Preface

This blog post continues the previous one 《 Get along well with CANN One stop solution for target detection and recognition 【 Introduction 】》, If yes CANN Students who don't understand the one-stop scheme of target detection and recognition , It is recommended to read the previous blog first .

know AclLite

AclLite What is it? ？

CANN The one-stop scheme of target detection and recognition is based on AclLite Developed , that AclLite And what is it ？ Simply speaking , It is further encapsulated based on Huawei shengteng's native interface and official interface documents , In particular, it encapsulates the relevant duplicate code in the current open source community sample , A set of simple public interfaces for users .
In a word , Is to simplify development , Reduce based on CANN Difficulty of development . Its function introduction and acquisition method are shown in the figure below ：

AclLite brief introduction

Considering the actual development needs and popular applications , Now let's introduce the common processing methods ：

The image processing

Video processing

Model reasoning , Common functions and resource management

Applications and multithreading

Apart from the above , In fact, there are other parts , This is not going to unfold one by one , It should be noted that , This is the encapsulation based on the official interface , It belongs to secondary development , It is not recommended to use it as a standard library , But if you want to achieve it quickly , Verify the prototype , It can still be considered .

CANN Analysis of one-stop solution for target detection and recognition

seen 《 Get along well with CANN One stop solution for target detection and recognition 【 Introduction 】》 My friends should know , The original scheme is to detect the input vehicle first , Then classify the detected vehicles by color , It's a test + Cascading tasks of classification . The following describes the processing flow of these two tasks .

Test pretreatment

Video detection preprocessing

Detect the reasoning process

Post test processing flow

Classification pretreatment process

Classification reasoning process

Post classification process

Customize your application

This scheme is to provide a sample for developers , It is convenient for everyone to develop their own applications , We hope to use this scheme as a reference or template , Can quickly realize the development of their own applications , The overall replacement process is shown in the figure below ：

Preprocessing

Reasoning process

Analyze the model to be replaced , Configure variables representing the path of the model file , Initialize with the replaced model AclLiteModel Class object , According to the actual needs of the replaced model , Replace the initialization model input interface , And modify the execution interface of the model .
Generally speaking , Replace the model , From preprocessing to model loading , Reasoning and post-processing all need to be changed , Basically, it's refactoring code , But if the pretreatment or post-processing of the model has little difference , Then the changes will be much smaller , such as YOLO The pretreatment and post-treatment of the series are almost the same , It'll be much better .

post-processing

In addition to the above basic processing flow , You can also do it according to your needs , Increase what we want , More flexible functions , Such as dynamic Batch, The specific code modification is shown in the figure below ：

Besides , There are also dynamic resolution and dynamic AIPP And so on , But generally speaking , Dynamic characteristics , Although more flexible , But it usually makes the efficiency lower , Personally think that , Unless you have to use it to complete the established function , Otherwise, static is better . After all, reasoning tasks , Efficiency or performance is the more important consideration .

Conclusion

Not much said , Take the last picture of the teacher PPT As an end ！