AI Vision software
AI-powered Vision Software
The Robovision AI Vision Software was developed based on extensive experience in large-scale and unpredictable production environments. After more than a decade of implementing computer vision AI solutions across the world, two main success criteria emerged: computer vision AI implementations require reliable data and should be easy to maintain.
The Robovision AI vision sofware tackles these challenges—and is user-friendly, too.
How our software works
1. Data Import
✓ Upload compressed folders automatically tagged with metadata
✓ Customizable importer tailored to specific company workflows
✓ Metadata like Process ID or Device ID included for seamless integration
✓ Easily searchable & filterable
2. Data Annotation
✓ Advanced annotation tools, including single and multiview labeling
✓ GrabCut and Magnetic Lasso for fast segmentation
✓ Filtering options based on user, date, or review status metadata
✓ Predictive Annotator for AI-assisted labeling of new data (with user adjustments)
✓ Confusion matrix for annotation comparison
3. Data Curation
Tagging and filtering based on metadata to organize and inspect data, plus
✓ Training data analytics ensure class balance with graphical class distribution analysis
✓ Auto training/validation set creation with stratified splitting for balanced class representation
✓ Ground truth selection supports flexible strategies: selecting last updated annotations, annotations by specific users, manual selection, or random sampling
✓ Immutable training datasets with fixed links to select ground truth annotations
4. Model Training
✓ Interactive progress monitor for real-time visualization of training metrics
✓ Supports hyperparameter tuning
✓ Configure to industry-standard parameters: early stopping, input resolution, batch size, learning rate, and more with Default or Expert mode
✓ Transfer learning capabilities to other users to improve existing models by training with additional datasets.
✓ Assess multiple models based on performance, parameter settings, and the datasets used with detailed trained model comparison
5. Model Testing
Evaluate new model performance against the ground truth.
✓ Robust model performance evaluation
✓ Assess models against annotated test sets or specific user inputs
✓ Supports test comparison; select the best model by comparing same dataset performance with multiple models
✓ Model self-check feature allows the model to test its validation data
✓ Identify outliers and potential label impurities for higher model accuracy and data quality
6. Model Optimization
Improve classification performance with optimized resource management.
✓ Powerful model optimization tools
✓ Can classify uncertain samples into an “unknown” class, protecting against model drift; triggers model maintenance when the “unknown” ratio becomes high
✓ Platform includes class confidence threshold optimization
✓ Set thresholds per class to identify valuable samples for re-labeling post-inference
✓ Model confidence threshold optimization adjust overall confidence levels based on available manual classification capacity
7. Model Deployment
Model performance and reliability in a variety of production environments.
✓ For running inference: flexible deployment options, centrally or on fab floor
✓ Configure parameters and choose deployment options
✓ With Model Inference feature, send new samples and receive predictions via API endpoint
✓ Model Monitor offers detailed reporting, tracking metrics like “unknown” rate and identifies low-confidence samples