AugeLab Studio Manual
English
English
  • 👋Welcome to AugeLab Studio User Manual
  • 📘Introduction
    • AugeLab Studio
    • Key Features
    • Use Cases
    • System Requirements
  • 🚀Getting Started
    • Signing up
    • Installation
    • First Look
    • Simple Tour
    • Your Very First Project
      • Basics
      • Detection
      • Wrapping Up
    • More Local Examples
    • Further Reading
  • 🖥️AugeLab Studio Interface
    • Detailed Look
    • Scenario Area
    • Menu and Toolbar
    • Managing Projects
    • Installing AI and much more
      • Leverage AI with Module Downloader
  • 🧱Function Blocks
    • Block Structures
    • Sockets
    • Blocks Column
    • Connections
    • All Function Blocks
      • AI Blocks
        • Face Detection
        • Mask Detection
        • Object Detection - Custom
        • Object Detection
        • Pose Estimation
        • Safety Equipment Detection
        • Social Distance Detector
        • Super Resolution
        • Text Detection
        • OCR
      • CNN Blocks
        • Average Pooling 2D
        • Batch Normalization
        • Choose Folder 2D
        • Compile Model
        • Conv. Sep. Layer 2D
        • Conv. Trans. Layer 2D
        • Convolutional Layer 2D
        • Dropout Layer
        • Flatten Layer 2D
        • Fully Connected
        • Global Average Pooling 2D
        • Global Max Pooling 2D
        • Input Layer 2D
        • Loss CCE
        • Max Pooling 2D
        • Metrics Accuracy
        • Model EfficientNet
        • Model MobileNet
        • Model ResNet
        • Model VGG
        • Optimizer Adadelta
        • Optimizer Adagrad
        • Optimizer Adam
        • Optimizer Adamax
        • Optimizer FTRL
        • Optimizer Nadam
        • Optimizer RMSProp
        • Optimizer SGD
        • ReLU Layer
        • Softmax Layer
        • Training Parameters
      • Data/Logic
        • Flow Control
          • Batch Concatenation
          • Batch Processing
          • Debatch
          • Get Batch Size
          • HMI Background
          • Subsystem Enabled
          • Subsystem In
          • Subsystem Loop
          • Subsystem Out
          • Subsystem
        • logic
          • All True
          • And
          • Demux
          • Equals
          • Greater
          • Logic Operations
          • Mux
          • Not
          • Or
          • Set - Reset
          • Smaller
        • Mathmetical Operations
          • Add
          • Counter
          • Divide
          • Math Operations
          • Maximum
          • Minimum
          • Multiply
          • Not Equals
          • Round
          • Square Root
          • Subtract
          • Trigonometry
        • Data Operations
          • Data Memory
          • Data to JSON
          • Data Type Converter
          • Datetime Compare
          • Dictionary Operations
          • Exclude Nones
          • Find Substring
          • Get Element
          • Is None
          • List Operations
          • Parse Data Dictionary
          • Replace None
          • String Merge
          • String Operations
        • Referencing
          • Data Read Global
          • Data Read Local
          • Data Write Global
          • Data Write Local
          • Debug Input
          • Tag From
          • Tag To
        • Signal Operators
          • Delay Step
          • Edge Falling
          • Edge Rising
          • Multi Port Switch
          • OFF Delay
          • ON Delay
      • Image/Transformations
        • Analysis
          • Color Density Percentage
          • Get Dimension
          • Histogram On Curve
          • Histogram On Line
          • Image Color Match
          • Image Memory
          • Image Resolution and Channel Value
          • Maximum Images
          • Mean Value of Image
          • Measure Position Distance
          • Minimum Images
          • Non-zero of Image
          • Std. of Image
          • Structural Similarity
        • Transformation Filters
          • Auto Alignment
          • Auto Contrast
          • Color Quantizer and Clustering
          • Color Space
          • Contrast-Brightness-Gamma
          • Contrast Optimization
          • Deconvolution
          • Denoising
          • Distance Transformation
          • FloodFill
          • Grab Cut Algorithm
        • Color Filters
          • 2D Filter
          • Apply Mask
          • Bilateral Filter
          • Blur
          • Edge Filter
          • HSV Filter
          • Image Adaptive Threshold
          • Image Threshold
          • Invert Image
          • Morphological Transformations
          • Normalize Image
          • RGB Mask
          • RGB Set
          • Sobel Filter
        • Operations
          • Add Images Weighted
          • Add Images
          • Collage Images
          • Divide Images
          • Flip Image
          • Image AutoRotator
          • Image Concatenate
          • Image Resize
          • Image Resizer
          • Merge Channels
          • Multiply Images
          • Polar Transform
          • Rotate Image Angle
          • Slice Image
          • Split Image
          • Subtract Images
      • Detections/Shapes
        • Detectors
          • Barcode Reader
          • Blob Detector
          • Blur Detector
          • Circle Detector
          • Corner Detector
          • Custom CNN Model
          • Data Matrix Reader
          • Detect Reference
          • Feature Detector
          • Find Object - Multiple Image
          • Find Object
          • Find Reference
          • Harris Corner Filter
          • Line Detector
          • Match Shapes
          • Measure Object Distance
          • Shape Detector
        • Draw
          • Draw Detections
          • Draw Line
          • Draw Point
          • Draw Rectangle
          • Draw Result On Image
          • Write Date On Image
          • Write Text On Image
        • Roi Processing
          • Check Area (Polygon)
          • Check Area
          • Get Pixel Mouse
          • Get Pixel
          • Get ROI
          • Image ROI Center
          • Image ROI Polygon
          • Image ROI Select Multi
          • Image ROI Select
          • Image ROI
          • Perspective Transform
          • Rectangles in Rectangle
        • Shape Analysis
          • Approximate Contour
          • Choose Line
          • Contour to Image
          • Fill Contour
          • Find Contour
          • Hull Convex
          • Minimum Circle
          • Minimum Ellipse
          • Minimum Rectangle
          • Minimum Rotated Rectangle
          • Most Similar Shape
          • Point Polygon Test
      • Input/Output
        • Communication
          • Modbus Connect
          • Modbus Read
          • Modbus Write
          • MQTT Publish
          • MQTT Subscribe
          • OPC UA Client
          • OPC UA Read
          • OPC UA Write
          • REST API - Get
          • REST API - Post
          • Send Mail
          • Siemens S7 Connect
          • Siemens S7 Read
          • Siemens S7 Write
        • Data Inputs
          • Date-Time List
          • Date-Time
          • Headless Check
          • Keyboard/Barcode Reader
          • Logic Input
          • Number Input
          • Number Range
          • PWM (Pulse Width Modulation)
          • Rising Edge
          • String Input
          • Text
        • Image Inputs
          • Camera IP (ONVIF)
          • Camera IP
          • Camera USB External
          • Camera USB Vidgear
          • Camera USB
          • Load Image From Path
          • Load Image
          • Make Image
          • Pixel
          • Video
        • Outputs/Exports
          • CSV Export
          • Cycle Timer
          • File/Folder Operations
          • GPU Statistics
          • Image Logger
          • Image Write
          • Led Output
          • Multi Image Write
          • Output
          • Scope
          • Show Image
          • Stop
  • 📡Devices and Communication
    • Camera Usage
    • Communication Protocols
    • Further Reading
  • 🧩Example Projects
    • Demo Projects
    • Circumference Measurement
    • Object Counting
    • Tile Width Measurement
    • Human Detection
    • Object Detection
  • 🔑Key Features
    • Deploy Custom HMI Applications
    • Annotate Data for Object Detection
    • Train Custom AI Models
      • Choosing the Right Database
      • When to Stop Training
    • Create Plugins
      • Components
      • Coding Reference
    • Share Your Solutions with Community
    • Instal Python Packages
  • 📑FAQ
    • Contact Us
    • FAQ
    • Setting up a full project
  • Additional Resources
    • Training Schedule
    • Training Materials
    • AugeLab Experts
  • Appendix
    • Dictionary
    • References
Powered by GitBook
On this page
  • Introduction
  • Define Detection Requirements
  • Consider Database Size
  • Bounding Box Quality
  • Data Balance and Bias
  • Augmentation

Was this helpful?

  1. Key Features
  2. Train Custom AI Models

Choosing the Right Database

PreviousTrain Custom AI ModelsNextWhen to Stop Training

Last updated 8 months ago

Was this helpful?

Introduction

The success of an object detection model heavily relies on the quality and diversity of the database used for training and evaluation.

Selecting the appropriate database is a critical step in the development process, as it directly impacts the model's performance and generalization capabilities.

Define Detection Requirements

Before selecting a database, it is essential to clearly define the requirements and goals of the object detection model. Consider the following factors:

  • Object Categories: Identify the specific types of objects the model needs to detect. The database should contain a representative sample of these categories.

  • Image Variability: Assess the variation in image quality, lighting conditions, and object poses present in the target application. Choose/create a database that captures these real-world scenarios.

  • Scale and Complexity: Determine the scale and complexity of objects. Some databases focus on small objects, while others cover larger or more complex scenes.

Consider Database Size

The size of the database directly affects the model's performance and training time. For models with a large number of parameters, a larger database is generally preferred. However, small databases might suffice for simpler models or proof-of-concept implementations.

A good rule of thumb is to use a minimum of 1000 images per class. This ensures each class can be taught with precision.

Bounding Box Quality

Ensuring good bounding box quality will ensure better learning process. Two examples of good and bad bounding box implementations:

You see, good annotation only encapsulates the object bounding box. This makes easier for model to be optimized and gives less importance on the background.

Data Balance and Bias

Ensure that the database is well-balanced across different object categories to avoid bias towards dominant classes.

A skewed distribution may lead to suboptimal performance on underrepresented categories during training and evaluation.

Dividing visually indistinguisable objects into different classes may yield an ineffective training process.

Augmentation

Augmentation is a method of artificially creating new data by modifiying the existing ones. In image processing, this is done by modifying constrast, values, width, height, perspective and many more feature to create a more generic database.

You can use Plug-in Window download augmentation tools in AugeLab Studio.

🔑