Quick Start - Semantic Segmentation Demo

Semantic segmentation is used to detect and provide pixel-level segmentation results for target objects in various scenes. It is also commonly used for defect detection, such as detecting cracks, defects, dirt, color variations, scratches, and spots in industrial environments.

The development of a Semantic segmentation model involves 7 steps: Create Model - Upload Images - Label Images - Train Model - Test Images - Export Model - Download SDK

The following example is about break detection using semantic segmentation. The purpose of this example is to identify and segment the break areas in the image.

The specific steps for model development are as follows:

1.Create Model

Click on "Create a model".

Enter the Model name (up to 20 characters), select the Model type as "Semantic Segmentation", and click "Confirm" to create the model.

If you want to use the Crop an image to several parts, click here for more information! Crop image

2.Upload Images

Click on "Dataset" to upload images. Click on "Upload Image".

Check the requirements for uploading images, click on "Upload now", select the images, and upload them.

Batch uploading is supported.

After the upload is complete, the upload results will be displayed, and you can see that all the images have been successfully uploaded. This completes the image uploading process.

If you want to upload locally labeled images,click here for more information! Upload img+LBL

3.Label Images

Enter the Labeling interface where you can label the images. The red box in the image represents the toolbar.

On the left side of the toolbar is the labeling area. The first tool is a brush that can be used to smear and annotation target region. The second tool is a polygon annotation tool that can be used to annotation target region by click on polygon vertices. In the middle of the toolbar, you can zoom in and out of the image (you can also use the mouse scroll wheel to zoom). On the right side of the toolbar, there are some advanced features.

The semantic segmentation annotation process mainly consists of three steps:

1. Add a label to the label list. In this example, there is only one label named "break".

2. Use the "Label" icon in the toolbar to label the image. This operation allows you to use polygons to label the target regions.

3. After completing the annotation of a specific target region, press the "F" key to save. The annotated region will then be displayed with the label "break".

For detailed label rules and examples, please refer to "More" -> "Label Examples".

4.Train Images

Enter the "Training" interface and click on "Start Training".

The following screen will appear, indicating that the task is in the queue. Please wait.

Then the model will start training, which will take approximately 10 minutes.

5.Test Images

Enter the Test page and click on "Start Testing" to perform model testing.

Once the testing is complete, the following message will appear, indicating that the testing is finished. If you want to view the testing results, click on "View Test Results".

The predicted results will be displayed on the original image.

If you want to obtain a large number of pre-labeled images but label few picture , click here for more information! Test result transfer

6.Export Model

Enter the download page, select "Current Model" and corresponding GPU, and click on "Export Model".

If you want know the Inference Time of using the model on your GPU, click here.

7.Download SDK

On the download page, select "C++ V1.0". Click on "Download SDK".