Unity Image Recognition Project

This project demonstrates a real-time image classification system built with Unity 6 (6000.0.58f2) and Unity Sentis (formerly Barracuda). It uses a deep learning model to classify input images directly within the Unity engine.

Features

• In-Engine Inference: Runs AI models natively in Unity without external dependencies using Sentis.
• GPU Acceleration: Utilizes GPU Compute for efficient model execution.
• Image Processing: Automatically converts and resizes Texture2D inputs to the required tensor shape (1, 3, 224, 224).
• Top Prediction: identifies the most likely class from the model's output.
• Target Search: Ability to check confidence for specific target objects (e.g. "goldfish").

Technical Details

• Model: MobileNetV2 (mobilenetv2-10.onnx)
• Input Resolution: 224x224 RGB
• Backend: Unity Sentis BackendType.GPUCompute

Requirements

• Unity Version: 6000.0.58f2 or later (should work with later unity 6 versions too)
• Packages:
  • com.unity.ai.inference (Sentis) v2.2.2

Setup and Usage

1. Clone the Repository:

   git clone <your-repo-url>

2. Open in Unity:
   • Open the project via Unity Hub.
   • Ensure the Unity version matches or is compatible.
3. Configuration:
   • The main logic is in ClassifyImage.cs.
   • Ensure a valid .onnx model is assigned to the Model Asset field.
   • Ensure synset.txt (ImageNet labels) is assigned to the Labels Field.
   • Assign an input image to the Input Image field in the inspector.
4. Run:
   • Enter Play Mode.
   • The classification result will be logged to the Console.

File Structure

• Assets/Scripts/ClassifyImage.cs: Core logic for loading model, processing input, and interpreting results.
• Assets/AI Models/: Contains the mobilenetv2-10.onnx model file.
• Assets/synset.txt: List of class labels (ImageNet).

References

• Unity Sentis Documentation
• ONNX Model Zoo - MobileNet