Object Detection
Data Preparation
Prepare Training Set Images Object detection models require a large number of defective images to detect objects or known types of defects. It is recommended that defective images make up more than 80% of the total images.
Prepare Test Set For the test set, prepare a large number of defective images as well, with defective images accounting for less than 20% of the total images.
TIP
- If the number of images exceeds 200, please compress them into a zip file, and ensure the file size is under 5GB.
Create a Project
Log in to AI Creator, and click the "New" button in the project center.
Click on "Custom Project", enter the project name "Food Packaging - Liquor Bottle Cap Detection" in the pop-up window, and then click "Next: Import Project Data" to complete project creation.
Import Data
Click "Dataset" in the left-hand navigation bar.
Upload training and test image sets
- When the project is created, folders for the dataset and training data will be automatically generated. Navigate into either the training or test set folder and click "Import Images" -> "New Dataset". Drag and drop images or compressed files into the upload area, then click "OK" to import the data.
Add Algorithm Module
Click "Label & Train" in the left-hand navigation bar.
Click "Add Module", then choose "Localization" in the module selection window.
Link Images
Click the "Link Data" drop-down menu and select "Link Images". In the pop-up window, choose the training set, then click "OK" to confirm.
Data Labeling
Click "Label & Train" in the left-hand navigation bar.
Label Data Use the labeling tool (rectangle tool) to mark defective areas and save the label information for the entire image (shortcut: Ctrl+Shift+S).
TIP
- If there are multiple areas to be localized, label all required regions before saving the label information.
Model Training
Click the "Train Model" button in the top-right corner of the page to enter the model training settings.
Configure general parameters, such as input size, number of training epochs, learning rate, and batch size based on your needs.
For the balance between speed and accuracy, it is recommended to use a middle value.
In "Dataset Split", set the training set ratio to 80%.
For fresh training, select "Train from Scratch". To continue training from the previous model, choose "Add Sample Training".
Expand the test parameters and leave the confidence threshold and NMS threshold as default.
After setting the parameters, click "Start Training".
TIP
For detailed explanations of specific parameters, hover over the help icon next to the parameter.
Training Process
Once training begins, the page will automatically display the training details. You can view real-time logs, such as the current epoch, as well as monitor metrics like train/valid Loss, train/valid IoU.
Training Completion
AI Creator will automatically evaluate and save multiple trained models. After training is complete, the system will select the best-performing model and archive it as a version. The model will then undergo an initial evaluation based on this version.
Model Evaluation
Once training is complete, the system will automatically select the best model for evaluation and calculate performance metrics.
AI Creator allows evaluation of the training images, displaying results such as annotations, outcomes, and ROIs.
If the evaluation results do not meet expectations, iterate the following steps until satisfactory results are achieved:
- Continue training with additional samples: Select "Add Sample Training" and retrain.
- Add new images, annotate them, and then select "Add Sample Training" to retrain.
Model Validation
Once the evaluation meets expectations, independently validate the model using test set images.
Click "Project Validation" in the left-hand navigation bar.
In the model validation page, select test set images and click "Start Validation".
View the validation results. After validation is complete, the system will automatically navigate to the validation results page.
Model Optimization
- Click "Model Optimization" in the left-hand navigation bar.
- To run models trained on x86 platforms on ARM-based edge devices, model conversion and optimization are required. AI Creator supports both auto optimization and advanced optimization. For localization models, you can use either option. Select QONN for the platform and QCS6490 for the chip.
Deployment
Click "Deployment" in the left-hand navigation bar to access the deployment page.
Generate Smart Camera Application
In the deployment page, click the "Generate Smart Camera Application" button. In the pop-up window, enter the application ID, version number, name, and description. Select the model version to deploy and choose the data (static images used by the smart camera for inference), then click "OK" to generate the application.
Deploy Application
In the deployment page, click the "Deploy" tab to access the application deployment page.
Node Integration
Before deploying a model or application, the target device node must be integrated.
- Click the "Node Management" tab at the top of the page to enter the node management page.
- Click the "Integrate Node" button to open the node integration pop-up.
- In the pop-up, select "IP Integration", choose "Smart Camera" in the node type, and input the node name and IP address. Click "OK" to confirm.
After the device is integrated, the "Node Type & Connection Status" will show as "Smart Camera: Online".
Return to the Project Deployment Page
Click "Project Center" and, in the action list of the current project, click "Deploy" to enter the main deployment page.
Deploy Smart Camera Application
In the deployment page, click the "Deploy" tab, then in the actions list of the created smart camera node, click "Deploy Application" -> "Deploy".
Validate Deployed Application
After the application deployment is completed on AI Creator, run the deployed application on the smart camera node and check the results.
TIP
This part is not an AI Creator feature and is only a direction for the next steps.