type
status
date
slug
category
password
tags

1. 🎯 Objectives

  • Evaluate object localization and classification accuracy.
  • Highlight missed detections and misclassified labels.
  • Generate annotated images and JSONs for easy debugging.
  • Provide a standalone executable and also allow CLI-based usage for flexibility.

2. 🧩 Logic Structure (Folder & Code Overview)

3. ⚙️ Compile as a Single Executable

To compile a self-contained EXE with no external DLLs needed:
✅ Output: ConsoleApp1.exe (runnable standalone on any Windows 64-bit machine)

4. ▶ Run the App

✅ From Terminal (without compiling):

✅ From PowerShell (with compiled EXE):

💡 Example:

5. 📥 Inputs & 📤 Outputs

✅ Inputs

  1. Images Folder: .jpg images and matching _ground.json
  1. Localizer Model: localizer.onnx
  1. Classifier Archive: classifier.cat → contains:
      • classificationNet.onnx
      • labels.json
      • meanStd.json

✍️ Ground Truth Preparation Steps

  1. Run app to generate predicted outputs.
  1. Use makesense.ai or LabelImg to adjust annotations.
  1. Save the corrected file as imageName_ground.json.
💡 Ground Truth Format (per file):

✅ Outputs

  • predicted/test1_predicted.json: Detected boxes and labels
  • predicted/test1_boxChecker.jpg: Image annotated with mismatches
  • Console accuracy summary:

    6. 🧪 Parameters & Logic

    🔑 Passable Parameters

    • IoUThresh (default: 0.7) → Box matching threshold
    • angleThreshold, downsizeFactor, bayShlvCount
    • localizationThresh, displayThresh, deviation, maxDistanceCfnt, ...
    👉 Defined and validated in: OdnnRequestInfo.cs

    🧠 Logic

    1. Parameters are passed via command line.
    1. OdnnRequestInfo.cs checks:
        • Is key recognized (_knownParameters)?
        • Is value within valid range?
    🔍 This avoids runtime crashes and ensures consistent behavior.
    Big Data: SQL, Hive and ElasticSearchSmart Cropping Logic for Turntable Video Selected Frame Images