Overview

Inspector provides an end-to-end workflow for computer vision projects:

1. Upload - Import images or videos for your dataset
2. Label - Annotate objects in your images with bounding boxes or segmentation masks
3. Train - Train AI models using state-of-the-art algorithms
4. Test - Evaluate model performance on new data
5. Deploy - Deploy models to edge devices for real-time inference

Dashboard

The Dashboard provides a quick overview of your Inspector workspace, including project counts, computer status, and system resource usage. It serves as the central hub for monitoring your AI operations.

Dashboard Components

Component	Description
Project Count	Total number of projects in your workspace
Computer Count	Number of connected edge AI computers
Camera Count	Total cameras connected across all computers
GPU/VRAM/CPU/RAM Usage	Real-time system resource monitoring
Active Training Models	Models currently being trained
Deployed Models	Models deployed to edge devices

Projects

Projects are the core organizational unit in Inspector. Each project contains a dataset and associated models for a specific computer vision task. The Projects page displays all your projects as cards with key information including project type, image count, and labeled image count.

Project Types

Creating a New Project

To create a new project:

1. Click the CREATE PROJECT button on the right side of the Projects page
2. Enter a Name for your project
3. Add an optional Description
4. Select the Project Type (Object Detection, Object Segmentation, or Visual Anomaly Detection)
5. Click Create to create the project

Upload Images/Videos

After creating a project, you will be taken to the Upload interface where you can add images and videos to your dataset. This is the first step in the project workflow (Upload > Label > Train > Test). The Upload interface provides multiple methods to add data to your project.

Upload Interface Overview

The Upload interface consists of a large drag-and-drop area in the center and three action buttons at the bottom. You can also access this interface at any time by clicking the Upload step in the project workflow bar or by clicking the upload icon in the image list toolbar.

Element	Description
Drag-and-Drop Area	The large central area with a folder icon. Drag files directly from your computer and drop them here to upload. The icon animates when files are dragged over it.
Import dataset	Import a pre-packaged dataset in ZIP format. Useful for loading existing labeled datasets or transferring projects.
Choose Files	Open a file browser to select individual images or videos from your computer. Supports multiple file selection.
From Camera	Capture images directly from connected edge AI cameras. Redirects to the Data Collection interface.

Supported File Formats

Inspector supports a variety of image and video formats for maximum flexibility:

Category	Supported Formats	Notes
Images	PNG, JPG, JPEG, BMP	Standard image formats. PNG recommended for lossless quality.
Videos	MP4, MOV, AVI, MKV	Videos are automatically extracted into individual frames.
Datasets	ZIP	Compressed archives containing images and optional annotation files.

Upload Method 1: Drag and Drop

The fastest way to upload files is by dragging them directly onto the upload area:

1. Open your file explorer and navigate to the folder containing your images or videos
2. Select the files you want to upload (use Ctrl+Click or Shift+Click for multiple selection)
3. Drag the selected files over the upload area - the folder icon will animate to indicate it's ready to receive files
4. Drop the files to start the upload process
5. Wait for the upload to complete - a progress indicator will show the status

Upload Method 2: Import Dataset

Use this option to import a complete dataset packaged as a ZIP file. This is useful for:

• Importing datasets exported from other Inspector projects
• Loading pre-labeled datasets in standard formats (COCO, YOLO, etc.)
• Transferring projects between Inspector instances
• Bulk importing large numbers of images efficiently

To import a dataset:

1. Click the Import dataset button
2. Select a ZIP file from your computer
3. Wait for the import process to complete - Inspector will extract and process all files

Upload Method 3: Choose Files

Use this option to manually select files from your computer:

1. Click the Choose Files button
2. A file browser dialog will open
3. Navigate to the folder containing your images or videos
4. Select one or more files (hold Ctrl or Shift for multiple selection)
5. Click Open to start the upload

Upload Method 4: From Camera (Data Collection)

Capture images directly from connected edge AI cameras. Clicking the From Camera button redirects you to the Data Collection interface - a powerful 4-step wizard that allows you to configure automated image capture from your edge devices. This method is ideal for collecting real-world data directly from your production environment.

Data Collection Wizard Overview

The Data Collection wizard guides you through four steps to configure automated image capture from edge AI cameras. Each step must be completed before proceeding to the next. The wizard header shows your progress with checkmarks indicating completed steps.

Step	Name	Purpose
1	Information	Enter a name and optional description for this data collection session
2	Camera	Select an edge AI computer and choose a camera connected to it
3	Input	Configure input/output settings and select the target project for collected images
4	Extra Settings	Set image count, capture interval, resolution, and Inspector IP address

Step 1: Information

The first step requires you to provide basic information about this data collection session.

Field	Required	Description
Name	Yes	A unique identifier for this data collection session. Must be at least 3 characters. Example: "Camera_1", "Production_Line_A"
Description	No	Optional text to describe the purpose or context of this data collection. Useful for organizing multiple collection sessions.

After entering the name (and optionally a description), click Next to proceed to Step 2.

Step 2: Camera

In this step, you select which edge AI computer and camera will be used for data collection. All registered computers are displayed as expandable cards.

Element	Description
Computer Card	Each registered edge AI computer is shown as a card with its name (e.g., "Edge AI PC", "Jetson Edge AI PC") and IP address (e.g., 192.168.0.104)
Camera(s) Count	Shows the number of cameras connected to each computer
Refresh Icon	Click to reload the camera list from the edge device
Connected Camera(s)	Expandable list showing camera name, type (IP Camera, RVC2, SOLO), and a live preview thumbnail
Image Preview	Live thumbnail from the camera feed to help you identify the correct camera

Click on a computer card to expand it and view connected cameras. Then click on a camera row to select it (highlighted with a blue border). Click Next to proceed.

Step 3: Input

Configure input/output settings and select which project will receive the collected images. This step allows you to optionally configure external triggers for data collection.

Field	Options	Description
Input/Output Settings	None, MQTT, Siemens S7, Modbus TCP	Select an external trigger source for automated data collection. Choose "None" for manual/timed collection without external triggers.
Project	Dropdown list of all projects	Select the target project where collected images will be automatically added. All your Inspector projects are listed here.

Select the target project from the dropdown and click Next to proceed to the final step.

Step 4: Extra Settings

Configure the final parameters for data collection including how many images to capture, the capture interval, resolution, and the Inspector server IP address.

Parameter	Default	Description
Image Count	25	Total number of images to capture in this session. Only shown when Input/Output Settings is "None".
Interval (ms)	1000	Time between captures in milliseconds. 1000ms = 1 second. Lower values capture more frequently.
Inspector IP Address	Auto-detected	The IP address of the Inspector server where images will be sent. Usually auto-filled with your server's IP.
Resolution	1280x720 (720p)	The resolution at which images will be captured from the camera.

Starting Data Collection

After configuring all settings, click the Collect button to start the data collection process. The system will:

1. Connect to the selected edge AI computer
2. Start capturing images from the selected camera at the specified interval
3. Transfer each captured image to the Inspector server
4. Automatically add images to the selected project
5. Stop after reaching the specified image count (or continue indefinitely if using external triggers)

During collection, a loading indicator shows that the process is active. Once complete, you will be redirected to the Computers page where you can see the collection status.

Video Upload and Frame Extraction

When you upload video files (MP4, MOV, AVI, MKV), Inspector automatically extracts individual frames for labeling. A Frame Skip dialog will appear allowing you to control how frames are extracted:

Upload Limits and Batch Processing

Inspector handles large uploads efficiently through batch processing:

• Files are uploaded in batches of up to 900 files at a time
• Large uploads are automatically split into multiple batches
• Progress is tracked across all batches
• You can continue working while uploads process in the background

Image List (Dataset Gallery)

After uploading images or when opening an existing project, you will see the Image List interface. This is the central hub for managing your dataset - viewing uploaded images, selecting images for operations, filtering by label status, and accessing trained models. The Image List provides a visual gallery of all images in your project with powerful tools for dataset management.

Image List Interface Layout

The Image List interface consists of several key components designed to help you efficiently manage your dataset:

Component	Location	Description
Project Name	Top Left	Displays the current project name with a back arrow to return to the Projects list. Hover to reveal an edit icon for renaming the project.
Workflow Bar	Top Center	Shows the project workflow steps (Upload, Label, Train, Test) with checkmarks indicating completed steps. Click any step to navigate directly to that stage.
Image Count	Below Project Name	Shows the current page range and total image count (e.g., "1-20 of 50 images"). Click the number dropdown to change images per page (20, 50, 100, 500, 1000).
Toolbar	Top Right	Contains action buttons for Export, Drop Duplicates, Upload, AI Model Selection, and Filter.
Image Grid	Center	Displays image thumbnails in a responsive grid. Each image has a selection checkbox and optional status indicators.
Pagination	Bottom Center	Navigate between pages of images using numbered buttons or previous/next arrows.
MODELS Sidebar	Right Edge	Collapsible panel showing trained models for this project with their performance metrics.

Toolbar Actions

The toolbar at the top right of the Image List provides quick access to important dataset management functions:

Icon	Action	Description
Export (Arrow Up)	Export Dataset	Export your dataset as a ZIP file. You can choose the export format (COCO, YOLO, etc.) and include annotations. Useful for backup or transferring to other systems.
Search (Magnifying Glass)	Drop Duplicates	Automatically detect and remove duplicate images from your dataset. This helps clean up datasets that may have accidentally received the same images multiple times.
Cloud Upload	Upload Images/Videos	Opens the Upload dialog to add more images or videos to your project. Same interface as the initial upload after project creation.
Robot (SmartToy)	AI Model Selection	Select an AI model for assisted labeling. Opens a menu with Zero-Shot Models, Pre-trained Models, and Trained Models. The selected model will be used when you press 'L' in the labeling interface.
Filter (Funnel)	Filter Images	Filter which images are displayed based on their labeling status or class assignments. A badge shows the number of active filters.

Filtering Images

The Filter feature allows you to focus on specific subsets of your dataset. Click the Filter button to open the filter menu:

Filter Option	Description
Labeled	Show only images that have at least one annotation (bounding box or polygon)
Unlabeled	Show only images without any annotations - useful for finding images that still need labeling
Test set	Show only images marked as part of the test set (used for model evaluation, not training)
Class Names	Filter by specific classes - show only images containing annotations of the selected class(es)

You can combine multiple filters - for example, select both "Labeled" and a specific class to see only labeled images containing that class. The filter badge on the toolbar shows how many filters are currently active.

Image Selection and Actions

You can select images for bulk operations using several methods:

• Click Checkbox: Click the checkbox icon in the top-left corner of any image thumbnail to select/deselect it
• Click Image (when selected): When images are already selected, clicking another image toggles its selection
• Drag Selection: Click and drag on the background to create a selection rectangle - all images within the rectangle will be selected
• Select All: When images are selected, a "Select All" button appears to select all images on the current page
• Deselect All: Click "Deselect All" to clear your selection

When images are selected, additional options appear in the toolbar:

• Selection Count: Shows how many images are currently selected (e.g., "5 image(s) selected")
• Delete: Delete the selected images from your dataset (requires confirmation)

Image Status Indicators

Each image thumbnail displays visual indicators to help you understand its status:

Indicator	Location	Meaning
Checkbox (Circle with Check)	Top Left	Blue when selected, dark gray when not selected. Click to toggle selection.
Test Set Icon (T)	Top Right	Appears on hover or when image is in test set. Blue when in test set, gray otherwise. Click to toggle test set status.
"Unlabeled" Badge	Bottom Right	Displayed on images that have no annotations yet. Helps identify images that need labeling.

Opening Images for Labeling

To open an image in the labeling interface, simply click on the image thumbnail (when no images are selected). This will open the full labeling interface where you can:

• Draw bounding boxes (Object Detection) or polygons (Object Segmentation)
• Use AI Assist to automatically generate annotations
• Navigate between images using the carousel or keyboard shortcuts
• Manage classes and view annotation details

MODELS Sidebar

The MODELS tab on the right side of the screen provides quick access to all trained models for the current project. Click the tab to expand the sidebar:

Each model card in the sidebar displays:

• Model Name: The name you assigned during training
• Performance Metrics: F1 Score, mAP 0.5, mAP 0.9, Precision, Recall
• Configuration: Model Type (Tiny/Small/Medium/Large), Epochs, Batch Size, Image Size, Algorithm
• Actions Menu: Three-dot menu for additional options like viewing details or deleting the model

Click on a model card to navigate to the Training interface and view detailed results, or to use the model for testing.

Pagination and Images Per Page

For large datasets, images are displayed across multiple pages. You can control pagination using:

• Page Numbers: Click numbered buttons at the bottom to jump to a specific page
• Previous/Next Arrows: Navigate one page at a time
• Images Per Page Dropdown: Click the number next to the image count (default: 20) to change how many images are displayed per page. Options: 20, 50, 100, 500, 1000

After Upload: Next Steps

Once your images are uploaded, the project workflow guides you through the next steps:

1. For Object Detection/Segmentation projects: Proceed to the Label step to annotate objects in your images
2. For Visual Anomaly Detection projects: Proceed directly to the Train step (no labeling required - the model learns from "normal" images)

Labeling

The Labeling interface is where you annotate objects in your images to train AI models. This is a crucial step that directly impacts model accuracy. Inspector provides a powerful, intuitive labeling interface with support for bounding boxes (Object Detection), polygons (Object Segmentation), and AI-assisted annotation tools. This section covers everything you need to know to efficiently label your dataset.

Getting Started with Labeling

To access the labeling interface, navigate to your project and click the Label tab in the project workflow bar. The labeling interface will open with your first image displayed on the canvas.

Interface Layout

The labeling interface is divided into several key areas, each designed to help you work efficiently:

Component	Location	Description
Image Canvas	Center	The main workspace where you view images and draw annotations. Supports zoom (mouse wheel) and pan (drag in Pan mode).
Toolbar	Top	Contains mode buttons (Labeling, Pan, AI Assist), navigation controls, undo/redo, filter, and other tools.
Image Carousel	Bottom	Thumbnail strip showing all images in your dataset. Click to navigate between images.
Right Sidebar	Right	Contains three expandable panels: Hotkeys (H), Objects, and Classes. Also shows class quick-select buttons.
Project Name	Top Left	Displays the current project name as a watermark overlay.

Cursor Modes

The labeling interface has four cursor modes that determine how your mouse interacts with the canvas:

Object Detection Labeling (Bounding Boxes)

For Object Detection projects, you annotate objects by drawing rectangular bounding boxes around them. Here's how to create and manage bounding box annotations:

Creating a Bounding Box

1. Ensure you are in Labeling Mode (press W or click the crop icon in the toolbar)
2. Select the appropriate class from the right sidebar (use number keys 1-9 for quick selection)
3. Click and hold at one corner of the object you want to annotate
4. Drag to the opposite corner while holding the mouse button
5. Release the mouse button to complete the bounding box

Editing a Bounding Box

To modify an existing bounding box:

• Select - Click on a bounding box or select it from the Objects panel
• Resize - Drag the corner or edge handles to adjust size
• Move - Drag the center of the selected box to reposition it
• Change Class - With the box selected, click a different class in the sidebar or press 1-9
• Delete - Press Delete to remove the selected annotation

Object Segmentation Labeling (Polygons)

For Object Segmentation projects, you create pixel-precise polygon masks around objects. This provides more accurate boundaries than bounding boxes.

Creating a Polygon

1. Ensure you are in Labeling Mode (press W)
2. Select the appropriate class from the right sidebar
3. Click to place the first point of your polygon
4. Continue clicking to add more points around the object boundary
5. To close the polygon, click near the starting point (within 3 pixels)

AI-Assisted Polygon Labeling (SAM)

Inspector integrates with SAM (Segment Anything Model) to help you create polygon masks faster:

1. Press P to enter Polygon AI Assist mode
2. Left-click on points inside the object you want to segment (positive points)
3. Right-click on points outside the object to exclude areas (negative points)
4. The AI will automatically generate a polygon mask based on your clicks
5. Press N to reset and start a new polygon

Objects Panel

The Objects panel displays all annotations on the current image. Click the Objects tab on the right sidebar to expand it.

Feature	Description
Object List	Shows all annotations with their class, coordinates, and dimensions
Selection	Click an object to select it on the canvas
Visibility Toggle	Eye icon to show/hide individual annotations
Delete	Trash icon to delete individual annotations
Clear All	Button to remove all annotations from the current image
Track/Untrack	Mark annotations for tracking across video frames

Classes Panel

The Classes panel allows you to manage object classes for your project. Click the Classes tab on the right sidebar to expand it.

Class Management

• Create New Class - Click the "Create New Class" button to add a new class
• Select Class - Click a class or press 1-9 to select it for annotation
• Class Colors - Each class has a unique color for visual distinction
• Visibility - Toggle class visibility to show/hide all annotations of that class
• Hot Keys - Classes 1-9 can be quickly selected using number keys

Hotkeys Panel

The Hotkeys panel shows all available keyboard shortcuts and allows you to customize them. Click the H button on the right sidebar to expand it.

Complete Keyboard Shortcuts Reference

Master these keyboard shortcuts to significantly speed up your labeling workflow:

Action	Primary Shortcut	Alternative	Description
Labeling Mode	W	-	Switch to annotation drawing mode
Pan Mode	S	-	Switch to canvas panning mode
AI Assist	L	-	Trigger AI-assisted annotation
Polygon AI Assist	P	-	Enter SAM polygon assist mode (Segmentation only)
Reset Circle Points	N	-	Clear SAM click points and start new polygon
Undo	Ctrl+Z	Q	Undo the last action
Redo	Ctrl+Shift+Z	E	Redo the previously undone action
Previous Image	Arrow Left	A	Navigate to the previous image
Next Image	Arrow Right	D	Navigate to the next image
Save Image	Ctrl+S	-	Save current annotations
Unselect ROI	I	-	Deselect the currently selected annotation
Delete ROI	Delete	-	Delete the selected annotation
Duplicate ROI	Ctrl+D	-	Create a copy of the selected annotation
Copy ROI	Ctrl+C	-	Copy the selected annotation to clipboard
Paste ROI	Ctrl+V	-	Paste the copied annotation
Track/Untrack	T	-	Toggle tracking status for the selected annotation
Select Class 1-9	1 - 9	-	Quickly select a class by its number

Image Filtering

Use the Filter button in the toolbar to filter which images are displayed in the carousel:

• Labeled - Show only images that have annotations
• Unlabeled - Show only images without annotations
• Test Set - Show only images marked for testing
• By Class - Filter images containing specific classes

Canvas Controls

Navigate and zoom the canvas to work with images of any size:

• Zoom In/Out - Use the mouse scroll wheel to zoom
• Pan - In Pan mode (S), click and drag to move around the image
• Fit to Screen - The image automatically fits to the canvas when loaded
• Fullscreen - Click the fullscreen button in the toolbar to maximize the labeling interface

Best Practices for Labeling

Follow these tips to create high-quality annotations that will improve your model's accuracy:

1. Be Consistent - Apply the same labeling standards across all images
2. Tight Bounding Boxes - Draw boxes that closely fit the object without excessive padding
3. Label All Instances - Make sure to annotate every visible instance of each class
4. Handle Occlusion - For partially visible objects, annotate only the visible portion
5. Use Keyboard Shortcuts - Learn the hotkeys to significantly speed up your workflow
6. Save Frequently - Press Ctrl+S regularly to save your work
7. Review Your Work - Use the Objects panel to verify all annotations before moving to the next image

Toolbar Reference

The toolbar at the top of the labeling interface provides quick access to all labeling functions:

Button	Function	Hotkey
Crop Icon	Labeling Mode - Draw annotations	W
Hand Icon	Pan Mode - Navigate the canvas	S
Robot Icon	AI Assist - Auto-annotate with AI	L
Circle Check Icon	Polygon AI Assist (Segmentation only)	P
Filter Icon	Filter images by status or class	-
Left Arrow	Previous image	A or Arrow Left
Right Arrow	Next image	D or Arrow Right
Undo Icon	Undo last action	Ctrl+Z or Q
Redo Icon	Redo last undone action	Ctrl+Shift+Z or E
Trash Icon	Delete current image	-
Fullscreen Icon	Toggle fullscreen mode	-
Exit Icon	Exit labeling and return to project	-

AI Assist Mode - Comprehensive Guide

AI Assist Mode is one of Inspector's most powerful features, enabling you to dramatically speed up the labeling process by using AI models to automatically detect and annotate objects in your images. Instead of manually drawing every bounding box or polygon, you can leverage pre-trained models, zero-shot models, or your own trained models to generate annotations automatically. This section provides a complete guide to understanding and using AI Assist Mode effectively.

Types of AI Models for Assistance

Inspector supports three types of AI models for assisted labeling, each with different use cases:

Model Type	Description	Best For
Zero-Shot Models (YOLOE)	Pre-built models that can detect objects based on text descriptions without any training. Uses YOLOE (You Only Look Once - Everything) technology.	Quick labeling of common objects, starting new projects, when you don't have a trained model yet
Pre-trained Models	Models uploaded to Inspector that were trained on external datasets. These models have learned to detect specific object classes.	Specialized domains where you have access to pre-trained weights, transfer learning scenarios
Trained Models	Models you have trained within Inspector on your own labeled data. These are the most accurate for your specific use case.	Production labeling, iterative improvement, when you have already trained a model on similar data

Step 1: Selecting an AI Model for Assistance

Before you can use AI Assist Mode, you must first select which AI model will be used for automatic annotation. This is done from the project's image list page (Upload tab):

1. Navigate to your project and go to the Upload tab where your images are displayed
2. Look for the Robot Icon (SmartToy icon) in the toolbar at the top right of the image list
3. Click the Robot Icon to open the AI Model Selection Menu
4. The menu displays three categories of models:
   • Zero-Shot Models - YOLOE and other models that work without training
   • Pre-trained Models - Models you've uploaded to Inspector
   • Trained Models - Models you've trained within this project
5. Click on a model to select it
6. After selection, the Label Mapping Dialog will appear (see next step)

Once selected, the model name will be displayed below the toolbar showing "Selected AI Model For Assist: [Model Name]".

Step 2: Configuring Label Mapping

After selecting an AI model, you need to configure how the model's output labels map to your project's classes. This is crucial because the AI model may use different class names than your project.

The Label Mapping Dialog shows:

• Project Labels (left column) - The classes defined in your project
• Model Labels (right column) - Dropdown menus to select which model output maps to each project class
• Ignore Option - Select "Ignore" to exclude a project class from AI-assisted annotations

Configure the mapping carefully, then click Save to store your mapping. The mapping is saved per model, so you only need to configure it once for each model you use.

Step 3: Using AI Assist in the Labeling Interface

Once you have selected an AI model and configured label mapping, you can use AI Assist Mode in the labeling interface:

1. Open the labeling interface by clicking on any image in your project
2. Click the Robot Icon (SmartToy) in the toolbar, or press L
3. The AI Assist Threshold Dialog will appear
4. Enter a Threshold value between 0 and 1 (see threshold explanation below)
5. Click OK to run AI-assisted annotation
6. The AI model will analyze the current image and automatically create annotations
7. Review the generated annotations in the Objects panel on the right
8. Adjust, delete, or add annotations as needed
9. Press Ctrl+S to save your work

Understanding the Threshold Parameter

The threshold is a confidence value between 0 and 1 that controls how selective the AI is when creating annotations:

Threshold Value	Behavior	When to Use
Low (0.1 - 0.3)	More detections, including uncertain ones. May include false positives.	When you want to catch all possible objects and manually filter out incorrect ones. Good for recall-focused labeling.
Medium (0.4 - 0.6)	Balanced detection. Reasonable confidence with moderate false positives.	General-purpose labeling. Good starting point for most projects.
High (0.7 - 0.9)	Fewer detections, but higher confidence. May miss some objects.	When you want only high-confidence detections. Good for precision-focused labeling.
Very High (0.9 - 1.0)	Only the most confident detections. Will miss many objects.	When you need very precise annotations and prefer to manually add missed objects.

AI Assist for Object Detection (Bounding Boxes)

When using AI Assist Mode in an Object Detection project, the AI model will automatically create bounding box annotations around detected objects:

• Each detected object gets a rectangular bounding box
• The box is assigned to the appropriate class based on your label mapping
• Annotations appear in the Objects panel with coordinates (x, y, width, height)
• You can resize, move, or delete any annotation as needed
• Duplicate detections are automatically filtered out

AI Assist for Object Segmentation (Polygons)

When using AI Assist Mode in an Object Segmentation project, the AI model creates polygon masks that precisely outline detected objects:

• Each detected object gets a polygon mask following its contours
• Polygons provide pixel-level precision for object boundaries
• The polygon is assigned to the appropriate class based on your label mapping
• You can edit polygon points to refine the mask if needed
• Works with both trained models and zero-shot models (YOLOE)

Polygon AI Assist Mode (SAM)

In addition to the standard AI Assist Mode, Object Segmentation projects have access to Polygon AI Assist Mode, which uses SAM (Segment Anything Model) for interactive segmentation:

To use Polygon AI Assist Mode:

1. Press P or click the Circle Check Icon in the toolbar to enter Polygon AI Assist mode
2. Left-click on points inside the object you want to segment (positive points - shown in green)
3. Right-click on points outside the object to exclude areas (negative points - shown in red)
4. SAM will automatically generate a polygon mask based on your clicks
5. Continue adding points to refine the segmentation
6. Press N to reset and start a new polygon
7. Select the appropriate class for the generated polygon

Workflow: Efficient Labeling with AI Assist

Here's a recommended workflow for using AI Assist Mode to label your dataset efficiently:

1. Initial Setup
   • Create your project and define all classes
   • Upload your images
   • Select an AI model (start with Zero-Shot if you don't have a trained model)
   • Configure label mapping
2. First Pass - AI Annotation
   • Open the labeling interface
   • Press L to trigger AI Assist with threshold 0.5
   • Review the generated annotations
   • Delete false positives, adjust bounding boxes/polygons as needed
   • Manually add any missed objects
   • Save and move to the next image
3. Train Your Own Model
   • After labeling 50-100 images, train a model on your data
   • Select your trained model as the AI Assist model
   • Continue labeling with improved accuracy
4. Iterative Improvement
   • As you label more images, periodically retrain your model
   • Each iteration improves AI Assist accuracy
   • Eventually, AI Assist will require minimal corrections

AI Assist Keyboard Shortcuts

Shortcut	Action	Description
L	AI Assist	Opens the threshold dialog and triggers AI-assisted annotation on the current image
P	Polygon AI Assist	Enters SAM mode for interactive polygon segmentation (Segmentation projects only)
N	Reset Circle Points	Clears SAM click points and starts a new polygon
W	Labeling Mode	Returns to manual labeling mode

Troubleshooting AI Assist

Issue	Possible Cause	Solution
"You must select an AI model firstly" error	No AI model has been selected for assistance	Go to the Upload tab, click the Robot icon, and select a model
No annotations generated	Threshold too high or model doesn't recognize objects	Lower the threshold value or try a different model
Too many false positives	Threshold too low	Increase the threshold value (try 0.7 or higher)
Wrong class assignments	Label mapping not configured correctly	Re-select the model and update the label mapping
AI Assist is slow	Large image or complex model	Wait for processing to complete; consider using a smaller model
SAM not generating polygons	Not enough click points or points placed incorrectly	Add more positive points inside the object and negative points outside

Best Practices for AI Assist

1. Start with Zero-Shot - If you don't have a trained model, YOLOE zero-shot models can provide a good starting point for common objects
2. Configure Label Mapping Carefully - Incorrect mapping will result in annotations assigned to wrong classes
3. Use Appropriate Threshold - Start at 0.5 and adjust based on your model's performance
4. Always Review AI Annotations - AI Assist is meant to speed up labeling, not replace human review
5. Train Your Own Model - For best results, train a model on your specific data and use it for AI Assist
6. Iterate and Improve - Periodically retrain your model as you label more data
7. Use SAM for Complex Shapes - For objects with irregular boundaries, Polygon AI Assist (SAM) often produces better results than automatic detection
8. Save Frequently - Press Ctrl+S after reviewing and correcting AI-generated annotations

Training

Training is the process of teaching an AI model to recognize patterns in your labeled data. Inspector provides a powerful, flexible training interface that supports multiple algorithms, training modes, and advanced hyperparameters. This section covers everything you need to know to train high-quality AI models, from basic Express Mode training to advanced Standard Mode with custom hyperparameters.

Getting Started with Training

To access the training interface, navigate to your project and click the Train tab in the project workflow bar. The training interface is divided into three main sections:

Section	Description
Project	Displays project information including name, type, and dataset statistics (visual count, labeled count)
Model Information	Shows a summary of current training configuration: Model Type, Image Size, Epochs, Algorithm, Batch Size, and Training Mode
Model Settings	The main configuration panel where you set all training parameters

Training Modes

Inspector offers two training modes to accommodate different user needs and experience levels:

Fine-tuning Models

Inspector supports fine-tuning, which allows you to train a new model starting from a previously trained model's weights. This is useful when you want to improve an existing model with additional data or adapt it to slightly different conditions.

To enable fine-tuning:

1. Check the Fine-tune Model checkbox in the Model Settings panel
2. Select a previously trained model from the dropdown (only models trained on the same project are available)
3. Configure your training parameters as usual
4. Click TRAIN to start fine-tuning

Algorithms for Object Detection and Segmentation

For Object Detection and Object Segmentation projects, Inspector supports the YOLO family of algorithms and RT-DETR (Real-Time Detection Transformer). Each algorithm has different characteristics in terms of speed, accuracy, memory usage, and stability.

YOLO Family Algorithms

Algorithm	Description	Speed	Memory	Accuracy	Stability
YOLOv7	Proven and stable YOLO version. Best for production environments where stability is critical.	Good	Medium	Good	Excellent
YOLOv8 (Recommended)	Enhanced architecture and training. Offers the best balance of speed, accuracy, and stability.	Better	Medium	Better	Good
YOLOv11	Latest optimizations and improvements. Excellent for high-performance applications.	Excellent	Medium	Excellent	Good
YOLOv12	Cutting-edge YOLO with best accuracy. Use when maximum accuracy is required.	Excellent	Medium	Maximum	Fair

Transformer-based Algorithm

Algorithm	Description	Speed	Memory	Accuracy	Stability
RT-DETR	Real-Time Detection Transformer. Uses attention mechanisms for excellent accuracy, especially on complex scenes.	Good	High	Excellent	Good

Model Size

Model size determines the complexity and capacity of the neural network. Larger models can learn more complex patterns but require more memory and compute time.

Size	Parameters	Speed	Memory	Best For
Nano/Tiny	~3-6M	Very Fast	Very Low	Edge devices, real-time applications, resource-constrained environments
Small	~11M	Fast	Medium	Balanced performance, general-purpose applications
Medium	~26M	Medium	High	Higher accuracy requirements, server deployments
Large	~44M	Slower	Very High	Maximum accuracy, offline processing, powerful hardware

Resolution (Image Size)

Resolution determines the input image size for the model. Higher resolutions provide better accuracy for detecting small objects but require more memory and training time.

Resolution	Training Speed	Memory Usage	Best For
320x320	Fastest	Lowest	Quick experiments, large objects, limited GPU memory
480x480	Fast	Low	Medium-sized objects, balanced performance
640x640 (Recommended)	Medium	Medium	General-purpose, good balance of accuracy and speed
736x736	Slower	High	Smaller objects, higher accuracy requirements
1024x1024	Slowest	Highest	Very small objects, maximum accuracy, powerful GPUs

Basic Training Parameters

Parameter	Description	Default	Recommendation
Model Name	A unique name for your trained model	-	Use descriptive names like "product-detection-v1"
Epochs	Number of complete passes through the training dataset	100	Start with 100, increase if model is underfitting
Batch Size	Number of images processed in each training iteration	16	Reduce if you get out-of-memory errors, increase for faster training

Algorithms for Visual Anomaly Detection

Visual Anomaly Detection projects use specialized algorithms designed to learn from normal samples and detect anomalies. These algorithms do not require labeled defect images - they learn what "normal" looks like and flag anything different.

Anomaly Detection Algorithms

Algorithm	Description	Speed	Memory	Accuracy	Complexity
PaDiM	Patch Distribution Modeling. Statistical approach that models the distribution of image patches. Good interpretability with anomaly heatmaps.	Fast	Low	Good	Low
EfficientAD	Efficient Anomaly Detection. Optimized for speed and resource usage. Ideal for real-time applications and edge deployment.	Very Fast	Very Low	Good	Low
DINO/DINOv3 (Recommended)	Self-supervised Vision Transformer. Uses powerful pre-trained features for excellent anomaly detection accuracy.	Medium	High	Excellent	Medium

Starting Training

Once you have configured all your training parameters, click the TRAIN button to start training. The training process will begin and you can monitor progress in real-time.

Training Requirements

• Object Detection/Segmentation: Requires labeled images with bounding boxes or polygons
• Visual Anomaly Detection: Requires at least 15 labeled images (labeled as "normal")

Training Results and Statistics

After training completes, Inspector displays comprehensive training results including performance metrics, training curves, and sample predictions.

Object Detection/Segmentation Results

Metric	Description	Good Value
F1 Score	Harmonic mean of Precision and Recall. Balanced measure of model performance.	> 0.90
mAP 0.5	Mean Average Precision at 50% IoU threshold. Standard detection metric.	> 0.90
mAP 0.9	Mean Average Precision at 90% IoU threshold. Measures precise localization.	> 0.70
Precision	Ratio of correct positive predictions to total positive predictions.	> 0.90
Recall	Ratio of correct positive predictions to total actual positives.	> 0.90

Model Statistics Visualizations (Object Detection/Segmentation)

• F1_curve.png - F1 score vs confidence threshold curve
• P_curve.png - Precision vs confidence threshold curve
• R_curve.png - Recall vs confidence threshold curve
• PR_curve.png - Precision-Recall curve showing the trade-off
• confusion_matrix.png - Matrix showing true vs predicted classes
• results.png - Training metrics over epochs (loss, mAP, etc.)

Visual Anomaly Detection Results

Metric	Description
Throughput (img/s)	Number of images the model can process per second. Higher is better for real-time applications.
GPU Memory (MB)	Amount of GPU memory used during inference. Important for edge deployment.
Val Mean	Mean anomaly score on validation set. Lower values indicate better separation.
Val Std	Standard deviation of anomaly scores. Lower values indicate more consistent predictions.
Val P95	95th percentile of anomaly scores. Useful for setting detection thresholds.
Val P99	99th percentile of anomaly scores. Useful for setting conservative thresholds.

Model Statistics Visualizations (Anomaly Detection)

• score_distribution.png - Distribution of anomaly scores across the validation set
• score_boxplot.png - Box plot showing score statistics
• model_summary.png - Summary of model architecture and parameters
• training_loss.png - Training loss over epochs (EfficientAD only)
• sample_anomaly_maps.png - Sample heatmaps showing detected anomaly regions (PaDiM and EfficientAD)

Managing Trained Models

After training, your models are saved and can be accessed from the MODELS panel on the right side of the training interface.

Model Actions

• View Details - Click on a model card to view detailed training results and statistics
• Test Model - Click "Let's Test Trained AI Model!" to evaluate the model on new images
• Delete Model - Click "DELETE MODEL" to remove a trained model (this action cannot be undone)
• Fine-tune - Use the model as a starting point for training a new model

Training Best Practices

Testing

Evaluate your trained models by testing them on new images or video streams. The testing interface allows you to adjust detection thresholds and visualize model predictions.

Testing Options

• Upload Images/Videos - Test with local files (up to 10 images or 1 video)
• Edge Device Camera - Test with live camera feed from connected edge devices

Performance Metrics

Metric	Description	Good Value
mAP@0.5	Mean Average Precision at 50% IoU threshold	> 90%
Precision	Ratio of correct positive predictions	> 90%
Recall	Ratio of actual positives correctly identified	> 90%

Computers

The Computers page is your central hub for managing edge AI devices and their connected cameras. Edge AI computers are physical devices (such as NVIDIA Jetson, industrial PCs, or other AI-capable hardware) that run your trained models for real-time inference. This section provides a comprehensive guide to managing computers, cameras, and dataflows in Inspector.

Understanding the Computers Page Layout

The Computers page consists of two main areas: the computer cards on the left and the action panel on the right.

Computer Cards

Each registered edge AI computer is displayed as a card containing:

Element	Description
Computer Name	A user-defined identifier for the edge device (e.g., "Edge AI PC", "Jetson Edge AI PC")
IP Address	The network address of the device (e.g., 192.168.0.104)
Cameras Button	Click to expand and view all cameras connected to this computer
Dataflows Button	Click to expand and view all AI dataflows deployed on this computer
Reload Icon	Appears when expanded - click to refresh cameras and dataflows from the device
Three-dot Menu	Access Edit and Delete options for the computer

Right Sidebar Actions

The right sidebar contains two vertical tabs for adding new devices:

• ADD COMPUTER - Opens the panel to discover and register new edge AI computers on your network
• ADD IP CAMERA - Opens the panel to add a new IP camera to an existing computer

Adding a New Computer

To add a new edge AI computer to Inspector, click the ADD COMPUTER tab on the right sidebar. This opens the "Find a Computer" panel.

Network Discovery Process

1. Set IP Range - Enter the starting IP address (e.g., 192.168.0.1) and the ending octet (e.g., 255) to define the search range
2. Click Search - Click the search icon to scan the network for edge AI devices
3. Select Device - From the discovered devices, click "Select" on the device you want to add
4. Configure Name - Enter a descriptive name for the computer (minimum 3 characters)
5. Add Description - Optionally add a description to help identify the device's purpose
6. Click Add Computer - Save the computer to your Inspector workspace

Adding an IP Camera

To add a new IP camera to an existing computer, click the ADD IP CAMERA tab on the right sidebar. This opens the "Add an IP Camera" panel.

Camera Types

Inspector supports two types of cameras:

Camera Type	Description	URL Format
IP Camera	Network cameras that stream video over RTSP protocol. Most common type for industrial and surveillance applications.	rtsp://username:password@ip:port/path
Integrated Camera (RVC2/SOLO)	Cameras directly connected to edge AI devices like OAK-D or other DepthAI hardware.	IP address of the camera module

Adding an IP Camera - Step by Step

1. Select Camera Type - Choose "IP Camera" or "Integrated Camera" based on your hardware
2. Select Computer - Choose the edge AI computer that will receive the camera feed from the dropdown
3. Enter Camera Name - Provide a descriptive name (minimum 3 characters)
4. Add Description - Optionally describe the camera's location or purpose
5. Enter URL/IP - For IP cameras, enter the full RTSP URL; for integrated cameras, enter the camera's IP address
6. Click Add IP Camera - Save the camera configuration

Viewing Connected Cameras

Click the Cameras button on any computer card to expand and view all connected cameras. The expanded view shows a table with detailed information about each camera.

Camera List Columns

Column	Description
Connected Camera(s)	The camera name and optional description
Type	Camera type: "Ip Camera", "RVC2" (integrated), or "SOLO"
URL	The RTSP URL or IP address used to connect to the camera
Image	A live thumbnail preview from the camera feed. Hover over integrated cameras to access the "Edit" button for camera settings.

Camera Settings (Integrated Cameras Only)

For integrated cameras (RVC2 and SOLO types), Inspector provides a comprehensive Camera Settings dialog that allows you to fine-tune image capture parameters in real-time. This is essential for optimizing image quality for your specific AI detection tasks. To access Camera Settings, hover over the camera's image thumbnail in the Cameras list and click the "Edit" button that appears.

Camera Settings Dialog Overview

The Camera Settings dialog consists of three main areas:

• Live Preview (Left) - Shows a real-time video stream from the camera so you can immediately see the effect of your adjustments. The preview updates automatically as you change settings.
• Settings Panel (Right) - Contains sliders for all adjustable parameters. Each slider shows the parameter name and current value in parentheses (e.g., "Exposure (5000)").
• Action Buttons (Bottom) - Three buttons for managing settings: Reload stream, Load default values, and Update values.

Camera Settings Parameters

The following table describes all available camera settings parameters with their valid ranges and recommended use cases:

Parameter	Range	Default	Description
Exposure	1 - 33000	5000	Controls how long the camera sensor is exposed to light (in microseconds). Lower values result in darker images but reduce motion blur. Higher values brighten the image but may cause blur on moving objects. For industrial inspection with good lighting, values between 3000-8000 work well.
ISO	100 - 1600	800	Adjusts the sensor's light sensitivity. Lower ISO values (100-400) produce cleaner images with less noise but require more light. Higher ISO values (800-1600) work better in low light but introduce more image noise. For AI detection, lower noise is generally preferred.
Focus	0 - 255	130	Sets the lens focus position for cameras with adjustable focus. Lower values focus on closer objects, higher values focus on distant objects. The optimal value depends on the distance between the camera and the objects being inspected. Adjust until objects appear sharp in the preview.
White Balance	1000 - 12000	4000	Adjusts color temperature in Kelvin. Lower values (1000-3000) produce warmer (yellowish) tones, while higher values (6000-12000) produce cooler (bluish) tones. Set to match your lighting conditions: ~2700K for incandescent, ~4000K for fluorescent, ~5500K for daylight, ~6500K for cloudy conditions.
Brightness	-10 to +10	0	Adjusts the overall brightness of the image. Negative values darken the image, positive values brighten it. Use this for fine-tuning after setting Exposure and ISO. Keep close to 0 for most accurate color representation.
Saturation	-10 to +10	0	Controls color intensity. Negative values reduce color saturation (toward grayscale), positive values increase color vibrancy. For defect detection where color is important, slight positive values (1-3) can help. For texture-based detection, 0 or slightly negative may work better.
Contrast	-10 to +10	0	Adjusts the difference between light and dark areas. Higher contrast makes edges more defined but may lose detail in shadows and highlights. For AI detection, moderate contrast (0-3) usually provides the best balance between edge definition and detail preservation.
Sharpness	0 - 4	1	Enhances edge definition in the image. Higher values make edges appear sharper but can introduce artifacts. For AI detection, values of 1-2 are recommended. Avoid maximum sharpness as it may create false edges that confuse detection models.
Luma Denoise	0 - 4	1	Reduces luminance (brightness) noise in the image. Higher values reduce more noise but may blur fine details. Use higher values (2-3) when shooting in low light with high ISO. Keep at 1 for well-lit environments to preserve detail.
Chroma Denoise	0 - 4	1	Reduces color noise (random colored pixels) in the image. Higher values reduce color noise more aggressively. Color noise is more common in low-light conditions. Values of 1-2 work well for most scenarios without affecting color accuracy.
Rotation	0°, 90°, 180°, 270°	0°	Rotates the camera image by the specified angle. Use this when the camera is physically mounted at an angle. Select the rotation that makes the image appear correctly oriented for your use case. This is applied before any AI processing.

Action Buttons

The Camera Settings dialog provides three action buttons at the bottom:

Button	Description
Reload stream	Refreshes the live camera preview. Use this if the preview becomes unresponsive or shows a stale image. The stream will reconnect and display the current camera view.
Load default values	Resets all camera settings to their factory default values. This is useful if you've made many changes and want to start fresh, or if the current settings are producing poor image quality.
Update values	Applies the current slider settings to the camera. Changes are sent to the camera hardware and take effect immediately. The live preview will update to show the new settings. Always click this button after making adjustments to save your changes.

How to Adjust Camera Settings

Follow these steps to optimize camera settings for your AI detection task:

1. Open Camera Settings - Navigate to Computers, expand the Cameras section for your device, hover over the camera thumbnail, and click "Edit".
2. Wait for Preview - The live preview will load showing the current camera view. A loading indicator appears while connecting to the camera stream.
3. Adjust Exposure and ISO First - Start by setting the overall brightness. Increase Exposure for brighter images, or increase ISO if you need to keep Exposure low to reduce motion blur.
4. Set Focus - Adjust the Focus slider until objects at your target distance appear sharp in the preview.
5. Fine-tune White Balance - Adjust White Balance to correct any color cast from your lighting. Objects that should appear white or gray should look neutral.
6. Adjust Image Quality Settings - Fine-tune Brightness, Saturation, Contrast, and Sharpness based on your specific needs.
7. Apply Noise Reduction - If you see noise (graininess) in the image, increase Luma Denoise and Chroma Denoise values.
8. Set Rotation if Needed - If the camera is mounted at an angle, select the appropriate rotation to orient the image correctly.
9. Click Update values - Apply your settings to the camera. The preview will update to show the final result.

Troubleshooting Camera Settings

Issue	Possible Cause	Solution
Preview not loading	Camera stream connection issue	Click "Reload stream" button. If still not working, check camera connectivity and ensure the camera is powered on.
Image too dark	Low Exposure or ISO	Increase Exposure first (up to 15000-20000), then increase ISO if needed. Also check Brightness setting.
Image too bright/washed out	High Exposure or ISO	Decrease Exposure and/or ISO. Reduce Brightness if it's above 0.
Image appears blurry	Incorrect Focus or motion blur	Adjust Focus slider until objects are sharp. If objects are moving, reduce Exposure to minimize motion blur.
Colors look wrong	Incorrect White Balance	Adjust White Balance to match your lighting type. Use lower values for warm lighting, higher for cool/daylight.
Grainy/noisy image	High ISO or low light	Increase Luma Denoise and Chroma Denoise. If possible, add more lighting and reduce ISO.
Settings not saving	Forgot to click Update values	Always click "Update values" after making changes. Settings are not applied until this button is clicked.

Managing Dataflows

Click the Dataflows button on any computer card to view and manage AI dataflows deployed on that device. Dataflows are the running AI pipelines that process camera feeds and generate inference results.

Understanding Dataflows

A dataflow is a diagram that controls data processing from the moment it is received from a source (camera) to the moment it is published to external environments. Each dataflow runs as a separate operating system service, ensuring they don't affect each other in terms of performance and scalability.

Dataflow List Columns

Column	Description
Dataflows on This Device	Name and description of the dataflow
Input	Input connector type (MQTT, Siemens S7, Modbus TCP) or "-" if none
Output	Output connector type (MQTT, Siemens S7, Modbus TCP) or "-" if none
Status	Green circle = running, Gray circle = stopped
Dashboard Icon	Click to open the dataflow's live dashboard in a new tab
Three-dot Menu	Access Edit, Start/Stop, and Delete options

Dataflow Capacity

Each edge AI computer has a maximum number of dataflows it can run simultaneously, shown as "X / Y" (e.g., "1 / 5" means 1 dataflow is deployed out of a maximum of 5). This limit depends on the device's hardware capabilities.

Dataflow Actions

• Deploy an AI Model - Click the button at the bottom to create a new dataflow using the Deploy wizard
• Edit - Modify the dataflow configuration (available from three-dot menu)
• Start/Stop - Toggle the dataflow's running state
• Delete - Remove the dataflow (only available when stopped)
• View Dashboard - Open the live inference dashboard to see real-time results

Dataflow Details Panel

When you click on a dataflow in the list, the right panel shows detailed information about the AI model used in that dataflow, including the model name, description, and class configurations with their detection thresholds and tracking settings.

Editing and Deleting Computers

To edit or delete a computer, click the three-dot menu icon on the computer card:

• Edit - Opens a dialog to change the computer's name and description
• Delete - Removes the computer from Inspector (requires confirmation)

Best Practices for Computer Management

1. Use Descriptive Names - Name computers based on their location or purpose (e.g., "Assembly Line 1", "Warehouse Entry")
2. Document Camera Locations - Use the description field to note camera positions and coverage areas
3. Monitor Dataflow Status - Regularly check that critical dataflows are running
4. Balance Dataflow Load - Distribute dataflows across multiple computers to avoid overloading a single device
5. Test Camera Feeds - Verify camera connectivity before deploying AI models
6. Keep IP Addresses Static - Configure static IP addresses for edge devices to prevent connection issues

Deploy

Deploy trained models to edge devices for real-time inference. The Deploy wizard guides you through a 5-step process to configure and launch your AI pipeline on edge devices.

Deployment Steps

1. Camera - Select the camera source and configure resolution
2. AI Model - Choose the trained model to deploy
3. Trigger - Configure when inference should run
4. Output - Define what happens with detection results
5. Track & Count - Enable object tracking and counting features

Deployment Configuration

Setting	Description
Name	Identifier for the deployment
Description	Optional description of the deployment purpose
Resolution	Camera resolution (e.g., 1280x720)

Model Registry

View and manage all trained models in your workspace. The Model Registry provides a centralized view of all your trained and pre-trained models with their performance metrics and configurations.

Model Information

Each model card displays:

• Model Name - Identifier for the model
• Performance Metrics - F1 Score, mAP, Precision, Recall
• Training Configuration - Model Type, Epochs, Batch Size, Image Size, Algorithm

Pre-trained Models

You can also add pre-trained models to your registry using the Add Pre-trained Model button. This allows you to use models trained elsewhere or download community models.

Overview

Project Studio allows you to create complex AI pipelines by connecting different processing nodes. This visual approach makes it easy to design, test, and deploy sophisticated data processing workflows.

Key Features

• Visual Flowchart Editor - Drag-and-drop interface for building workflows
• Workstation Management - Organize and manage multiple edge devices
• Dataflow Templates - Pre-built templates for common use cases
• Real-time Monitoring - Monitor dataflow execution in real-time

Workstations

Workstations represent the edge devices where your dataflows will run. Each workstation can have multiple cameras and dataflows configured.

Workstation Features

• Device Registration - Add and configure edge AI computers
• Camera Management - Connect and configure IP cameras
• Status Monitoring - View device health and connectivity

Flowchart Editor

The Flowchart Editor is the core interface for building AI dataflows. It provides a visual canvas where you can add, connect, and configure processing nodes.

Node Types

Building a Dataflow

1. Drag nodes from the toolbar onto the canvas
2. Connect nodes by drawing links between ports
3. Configure each node by clicking on it
4. Save and deploy the dataflow

Dataflows

Dataflows are the executable pipelines created in the Flowchart Editor. They define how data flows through your AI processing pipeline.

Dataflow Management

• Create - Build new dataflows using the Flowchart Editor
• Edit - Modify existing dataflows
• Start/Stop - Control dataflow execution
• Monitor - View real-time execution status and logs

Overview

Predictor enables you to build and deploy predictive models for various use cases including demand forecasting, predictive maintenance, and anomaly detection in time series data.

Key Features

• Multiple Algorithms - Support for LightGBM, Prophet, and other ML algorithms
• Environment Management - Organize predictions by environment/location
• Feature Importance - Understand which features drive predictions
• Model Deployment - Deploy models for real-time predictions

Environments

Environments represent the physical or logical locations where predictions are made. Each environment can have multiple items to predict.

Environment Configuration

• Environment Name - Identifier for the prediction context
• Items - Individual entities to predict (e.g., machines, sensors)
• Data Sources - Connected data feeds for training and inference

Datasets

Datasets contain the historical time series data used to train prediction models.

Dataset Requirements

• Timestamp Column - Date/time for each data point
• Target Variable - The value you want to predict
• Feature Columns - Additional variables that may influence predictions

Data Preparation Tips

Training

Train prediction models using your historical data and selected algorithms.

Supported Algorithms

Training Configuration

Parameter	Description
Forecast Horizon	How far into the future to predict
Training Window	Amount of historical data to use
Seasonality	Periodic patterns in the data (daily, weekly, yearly)

Testing

Evaluate model performance using holdout data and various metrics.

Evaluation Metrics

Metric	Description
MAE	Mean Absolute Error - average prediction error
RMSE	Root Mean Square Error - penalizes large errors
MAPE	Mean Absolute Percentage Error - relative error

Feature Importance

After training, you can view feature importance to understand which variables have the most impact on predictions. This helps in:

• Identifying key drivers of the target variable
• Removing irrelevant features to improve model performance
• Gaining business insights from the data