[Experience sharing] [Scene reproduction project based on AI camera] Maixhub online model training--Ultraman recognition training process

walker2048

[Experience sharing] [Scene reproduction project based on AI camera] Maixhub online model training--Ultraman recognition training process [Copy link]

 

Preface

In the previous post, we briefly talked about V831 and demonstrated the effect of classification recognition. This time, let’s talk about how to use Maixhub to train a classification model and deploy it to the device step by step.

---

Register an account and log in

First, we open the official website of maixhub and register an account (those who already have an account can log in directly).

---

Create a project

After logging in, you can see a button for model training and trial on the page, click it.

On the next page, click the Create Project button and enter the relevant information. Select Image Classification as the project type (no need to obtain object coordinates).

---

Project interface description

In the following project interface, there is a commonly used menu bar on the left, a general process of model training on the right, and an information window of the data set and training records below the flowchart.

---

1. Create, upload, and organize datasets

1.1 Create a dataset

Click the Create Dataset button, fill in the relevant information and confirm. After creating, remember to select the dataset.

1.2 Data collection interface description

Then when collecting data, we need to upload relevant photos of the dataset and verification set and mark them.

1.3 Get Ultraman's photo

It is actually not easy to get dozens of pictures related to Ultraman. You need to go to a website (such as Bilibili) to take screenshots and upload them.

After uploading the picture, just click to start training.

1.4 View training results

Wait patiently for the training to complete or click on the corresponding record in the training record to query the results.

2. Deploy the model

2.1 Follow the official tutorial to set up Wi-Fi for the development board.

2.2 After confirming that the IP address is obtained on the development board, select model deployment (deploy) in the menu

2.3 Then scan the model QR code provided on the web page. Wait for the model to be downloaded and deployed. There will be a progress bar prompt when downloading the model.

3. Validate the model

Open the original screenshot on the computer and verify whether it can be recognized. The following is the recognized photo.

The following is the identification video

93fab0809274d85390f2ec464c2f91d7

If you use static photos, it is still possible to recognize them, but dynamic videos are a bit troublesome.

---

Summarize

The training and deployment of the AI recognition model of V831 are relatively simple. In actual testing, the accuracy of static image recognition is still acceptable. Compared with object recognition, this online training platform may not be suitable for identifying different individuals of similar objects. In actual use, it is found that the recognition rate of different Ultraman is not particularly high. When selecting image detection type projects, the training cannot obtain an effective model (recognition is too low), and training often fails. According to past experience, it should be caused by too few photos of the model and too low clarity. In fact, if you want to complete the function of Ultraman recognition, it will take more time.