Auto Labeling

Background

• Only datasets of image classification and object detection types support the auto labeling function.
• To enable Auto Labeling, add at least two types of labels to the dataset and add each type of the label to at least 5 objects.
• At least one unlabeled image must exist when you enable Auto Labeling.
• Before enabling Auto Labeling, ensure that no auto labeling task is in progress in the system.
• Check the image data used for labeling and ensure that no RGBA four-channel image exists in the image data. If four-channel images exist, the auto labeling task will fail. Therefore, delete the four-channel images from the dataset and then start the auto labeling task.

Auto Labeling

1. Log in to the ModelArts management console. In the left navigation pane, choose Data Management > Datasets. The Datasets page is displayed.
2. In the dataset list, select a dataset of the object detection or image classification type and click Auto Labeling in the Operation column to start an intelligent labeling job.
3. On the Enable Auto Labeling page, select Active learning or Pre-labeling. For details, see Table 1 and Table 2.

   Table 1 Active learning

   Parameter	Description
   Auto Labeling Type	Active learning: The system uses semi-supervised learning and hard example filtering to perform auto labeling, reducing manual labeling workload and helping you find hard examples.
   Algorithm Type	For a dataset of the image classification type, you need to set the following parameters: Fast: Use the labeled samples for training. Precise: Use labeled and unlabeled samples for semi-supervised training, which improves the model precision.

   Table 2 Pre-labeling

   Parameter	Description
   Auto Labeling Type	Pre-labeling: Select an AI application created on the AI Applications page. Ensure that the model type matches the dataset labeling type. After the pre-labeling is complete, if the labeling result complies with the standard labeling format defined by the platform, the system filters hard examples. This step does not affect the pre-labeling result.
   Model and Version	My AI Applications: You can select a model based on site requirements. Click the drop-down arrow on the left of the target AI application and select a proper version. For details about how to create an AI application, see Creating an AI Application.
   Specifications	In the drop-down list, you can select the node specifications supported by ModelArts.
   Compute Nodes	The default value is 1. You can select a value based on site requirements. The maximum value is 5.

   

   • For datasets of the object detection type, only rectangular boxes can be recognized and labeled when Active Learning is selected.
   • If there are too many auto labeling jobs in the system, the jobs may be queued. As a result, the jobs are always in the labeling state. The system will complete labeling jobs in sequence.
   Figure 1 Enabling auto labeling (image classification)
   
   Figure 2 Enabling auto labeling (object detection)
   
   Figure 3 Enabling auto labeling (pre-labeling)
   
4. After setting the parameters, click Submit to enable auto labeling.
5. In the dataset list, click a dataset name to go to the Dashboard page.
6. On the Dashboard page of the dataset, click Label in the upper right corner. The dataset details page is displayed.
7. On the dataset details page, click the To be Confirmed tab to view the auto labeling progress.
   You can also enable auto labeling or view the auto labeling history on this tab page.

   Figure 4 Labeling progress
   
8. After auto labeling is complete, all the labeled images are displayed on the To Be Confirmed page.
   • Datasets of the image classification type

     On the To Be Confirmed page, check whether labels are correct, select the correctly labeled images, and click Labeled to confirm the auto labeling results. The confirmed image will be categorized to the Labeled page.

     You can manually modify the labels of the images marked as hard examples based on site requirements. For details, see For datasets of the image classification type.

   • Datasets of the object detection type

     On the To Be Confirmed page, click images to view their labeling details and check whether labels and target bounding boxes are correct. For the correctly labeled images, click Labeled to confirm the auto labeling results. The confirmed image will be categorized to the Labeled page.

     You can manually modify the labels or target bounding boxes of the images marked as hard examples based on site requirements. For details, see For datasets of the object detection type.