Road Marking Training

Function Description: Manually draw and create a marking library on the point cloud to train models, which can then be used for automatic extraction of markings in road sign matching.

Steps

1.Training Process

1)Click the Road Marking Training button and select the template file.

2)Click Next to select an existing table or create a new one (the default table cannot be used). Here, create a new table named 'tests'.

3) Click "Next" to start creating the template. If you already have a DXF or SHP template file, you can click the **Import** button, select the corresponding DXF or SHP file, and click **Open** to import the template and rename it.

If multiple templates exist in the SHP file, you can choose whether to name the template according to the existing fields in the SHP. If no field is used, the default name is used.

4)If there are no DXF or SHP templates, you can click Add to manually create a template in the template creation window.

First, fill in the template name, and then draw the template outline on the point cloud. When drawing, it is necessary to strictly follow a **counterclockwise** direction.

Drag the **Rotation ** slider to adjust the template, as shown in the image below.

After adding multiple templates, close the template creation window. Click the **Train** button to start the training.

5)You can set the training parameters here. The specific parameter settings are as follows: **GPU**: As shown in the figure, you can specify the GPU to be used for training from the dropdown menu. This feature only supports training on a GPU. **Batch Size**: This refers to the number of data inputs during each training iteration. For example, if the batch size is set to 2, then two sets of data will be trained in the same batch. Theoretically, a larger batch size can lead to higher precision results; however, it is recommended not to exceed 8. The upper limit is constrained by the size of the GPU memory. Based on different GPU memory sizes, the suggested batch sizes are listed in the table below. When you see a message about **insufficient memory or GPU memory**, you can adjust this parameter downwards.

GPU Memory Size

	ecommended Batch Size Settings
8GB	2
10GB	6
11GB	6
12 GB and above	8

Maximum Epochs: The maximum number of training cycles. This parameter should be adjusted based on the trend of Loss and other evaluation metrics, with the optimal value set when the Loss stops decreasing. The default is set to 200.

Pre-trained Model: An optional feature. This allows you to select a pre-trained model and can be used in two scenarios:

If training is unexpectedly interrupted, there is no need to retrain from scratch; you can simply use the saved model file at the time of interruption as the pre-trained model for training.
If you already have a trained model and new labeled data becomes available, there is no need to merge the old data to retrain. You can continue training based on the existing model, which can save time effectively. However, the categories of the training data and the model name must remain consistent between the two training sessions.

6)Click Next to enter the training interface. Click Start to begin training. This process may take a long time, so please be patient.

7)After training is complete, you can evaluate the model's convergence by observing the curve changes in the road marking training window. Refer to Image Deep Learning Training.

2.Inference Process

1) Once the model training is finished, click the road sign matching button, and select the table that was chosen or created in step 2. Here, select the tests table.

2) Select the automatic matching mode.

Click two points with the mouse, and drag to determine the transverse range of the markings, automatically extracting the road marking vectors. As shown in the image below, click the left mouse button in sequential order to select the candidate areas.

The final extraction results are shown below.

Road Marking Training

Road Marking Training

Steps

results matching ""

No results matching ""