Generate a Model with Geti™ Software#

This guide walks you through the process of installing the Geti™ software, setting up a pallet defect detection project, training a model, and deploying it.

Prerequisites#

Installation Steps#

For detailed Geti™ software installation instructions, refer to the Geti™ Installer Documentation.

Note: The standard Geti™ software installation includes the following steps:

  1. Download the Geti™ software installer

  2. Extract the installer archive

  3. Prepare the system by creating necessary directories

  4. Run the software installer with appropriate system privileges

Please follow the official installation guide for the most up-to-date and accurate installation procedures.

Upon successful completion, you will see the installation success confirmation as shown below:

Geti™ Software Installation

Setting Up Your Project#

Step 1: Sign In to Geti™ Software#

Open https://<host_ip> in your browser, where <host_ip> is the IP address of the system where you installed the Geti™ software server. Sign in with the credential which was set during installation:

Sign In to Geti™ Software

Step 2: Access Geti™ Software Dashboard#

After successful authentication, you will see the Geti™ software dashboard:

Geti™ Software Dashboard

Step 3: Create a New Project#

Click on “Create New Project” to start a new pallet defect detection project:

Create New Project

For detailed information refer the tutorial: Geti™ Software - Project Creation

Step 4: Select Detection Task#

Select “Detection” and choose “Detection bounding box” as your annotation type:

Select Detection - Bounding Box

Step 5: Create Labels#

Define the labels for your defect detection task (e.g., “defect”, “box”, “shipping label”, etc.):

Create Labels

For detailed information refer the tutorial: Geti™ Software - Label Management

Data Annotation and Training#

A dataset is a collection of images and videos in your project, together with the annotations used to train models. The Dataset screen is where you upload media, monitor annotation progress, and launch the annotator. For more information on datasets, see Dataset Management.

To use the Geti™ software for dataset creation and annotation, see Building a Good Dataset.

Step 6: Upload Training Images#

Browse and upload your training dataset images:

Browse and Upload Images

After uploading, your project dashboard will display the uploaded images:

Pallet Defect Detection Dashboard

Step 7: Annotate Images Interactively and Train the Model#

Click on “Annotate Interactively” on the top right side of the dashboard. Begin annotating your images manually:

Annotate Images

After annotating a few frames, the Geti™ software will automatically start training the model. For more information on model training, see Model Training.

Annotate a minimum number of frames to trigger automatic model training within the Geti™ software. A real-time estimate of the remaining “Annotations Required” before training starts is displayed in the upper-right corner of the interface.

Note: By default, the Geti™ software uses MobileNetV2-ATSS as the model backbone for your detection task. For more control over your model training, you can explore the Advanced Guide section below to:

  • Change model backbone to different architectures

  • Configure custom training parameters

  • Apply model optimization techniques (FP16, INT8)

Step 8: Monitor Training Progress#

You can monitor the model training progress in real-time:

Model Training

Step 9: Improve Model Accuracy (Optional)#

Repeat the annotation process to improve model accuracy. More annotated data will lead to better model performance.

Step 10: Download Model#

Download Model#

Click on the download icon next to the FP16 or INT8 model. A zip folder containing model.bin and model.xml will be downloaded. Replace the existing model files in your deployment resources:

model.bin  <- Replace with downloaded version
model.xml  <- Replace with downloaded version

Alternatively, you can download the entire deployment folder and replace the existing deployment folder in your resources:

Deployment Dashboard

For an introduction to the Inference Pipeline concept, see Inference Pipeline.

Step 11: Deploy Model#

Navigate to Deployments and click Select model for deployment:

Select Deployment Package

In the “Select model for deployment” dialog:

  1. Choose your desired Architecture

  2. Select your Optimization level (FP16 or INT8)

  3. Click Download

The deployment package will be downloaded. Replace the existing deployment folder inside your resources with this new package.

Advanced Guide#

The Advanced Guide section allows you to fine-tune your model training with more control over model architecture, parameters, and optimization.

Model Backbone Change#

Change the model backbone from the default architecture to other architectures for your specific requirements. For a complete list of supported model architectures, refer to Geti™ Software - Supported Models Documentation.

  1. Click on Models from the left sidebar

  2. Select Train Model

  3. Click on Advanced Settings

  4. Select your desired model type from the available options:

    • YOLOX-Tiny: Lightweight model for edge devices

    • YOLOX-Small: Small model with better accuracy

    • Other available backbone architectures

  5. Click Start to begin training with your selected backbone

For detailed information, refer to the tutorial: Geti™ Software - Model Training and Optimization Advanced Model Training

Monitor your selected backbone training progress:

YOLOX-Tiny Model Training

Train Parameters#

Configure custom training parameters to optimize model performance based on your dataset and requirements. For detailed information on available training parameters and their configurations, refer to Training Parameters Documentation.

Common parameters include:

  • Learning rate

  • Batch size

  • Number of epochs

  • Optimizer settings

  • Augmentation options

Training Parameters

Model Optimization#

After training completes, optimize your model for different deployment scenarios using quantization techniques. Choose the optimization level that best suits your deployment environment:

  • FP16: Higher precision with good accuracy, requires more computational resources

  • INT8: Optimized for edge deployment, significantly reduces model size and latency

Click on Start Optimization to generate your optimized model:

Select Trained Model and Optimization

After optimization, proceed with downloading and deploying your model.

Next Steps#

  • Deploy the model to edge devices

  • Monitor model performance

  • Continuously improve accuracy by adding more annotated data

  • Retrain as needed with new data

Troubleshooting#

For installation issues, refer to the Geti™ Software Installation Guide.