Unity Inference Engine For On-Device ML/CV Models

• Recompile and load the YOLO sample with Unity.
• Build a new project using Unity and Inference Engine.
• Compile and build Inference Engine models to support models other than YOLO.
• Integrate this API with the Unity Inference Engine architecture to access ML/CV models. Note that the Unity Inference Engine integration is not a required dependency and many developers will choose to access different or proprietary frameworks. However, it is provided as a convenience for developers to experiment quickly.

• Development machine running one of the following:
  • Windows 10+ (64-bit)
  • macOS 10.10+ (x86 or ARM)

• Supported Meta Quest headsets:
  • Quest 3
  • Quest 3S and 3S Xbox Edition

• Horizon OS v74 or later installed on your headset

• Unity Editor 2022.3.15f1 or later (6.1 or later recommended)

• Inference Engine package 2.2.1 (com.unity.ai.inference) installed in your project. See the Upgrade to Inference Engine⁠ if you used Unity Sentis before.
• Install Mixed Reality Utility Kit (MRUK) 85.0.0 or later.

• Grant Camera and Spatial data permissions to your app. The Spatial data permission is used by the EnvironmentRaycastManager.

1. The model accuracy is not 100% because the model has been optimized to work on devices like Quest 3.
2. This model has been trained with 80 classes (not objects), it means that some object will be included inside a class. For example: Monitor and TV are in the same class TV_Monitor. The table in the next section identifies the classes the model can identify.
3. Some classes are hard to identify, for example, cell phones are difficult to identify, in most cases are identified as the TV_Monitor class.
4. Quest 3 controllers might be unrecognized or be mislabeled as remote controllers.
5. The bounding boxes visualization in the MultiObjectDetection sample doesn’t perfectly align with the detected objects. For a better example of the camera to world projection, refer to the CameraToWorld sample.

1. Clone the Unity-PassthroughCameraApiSamples⁠ GitHub repository.
2. Open the Unity-PassthroughCameraApiSamples project in Unity Editor.
3. In Build Profiles switch the Active Platform to Android.
4. Open the Assets/PassthroughCameraApiSamples/MultiObjectDetection/MultiObjectDetection.unity sample scene.
5. Navigate to Meta > Tools > Project Setup Tool. If the rule “MR Utility Kit recommends Scene Support to be set to Required” exists, select ... > Ignore. Click Fix all and Apply all to resolve the other issues and recommendations.
6. Build the app and test it on your headset.

• Description: This sample shows how to identify multiple objects using the Inference Engine and a pretrained YOLO model.
• Controls: This sample uses the Quest 3 controllers:
  • Menus (Start and Pause):
    • Button A: start playing
  • In Game:
    • Button A: place a marker in the world position for each detected object
  • At any moment:
    • Button MENU: back to Samples selection.
• How to play:
  • Start the application and use the device to look around you.
  • When an object is detected, you will see 2D floating boxes around the detected objects.
  • If you press the A button, a 3D marker will be placed in the real world position of the detected objects with the class name.
  • This model can identify the following objects (80):
    People and animals:person, bird, cat, dog, horse, sheep, cow, elephant, bear, zebra, giraffe
    Vehicles and transport:bicycle, car, motorbike, aeroplane, bus, train, truck, boat
    Outdoor objects:traffic light, fire hydrant, stop sign, parking meter, bench
    Sports and accessories:frisbee, skis, snowboard, sports ball, kite, baseball bat, baseball glove, skateboard, surfboard, tennis racket
    Personal items:backpack, umbrella, handbag, tie, suitcase
    Kitchen and dining:bottle, wine glass, cup, fork, knife, spoon, bowl
    Food:banana, apple, sandwich, orange, broccoli, carrot, hot dog, pizza, donut, cake
    Furniture:chair, sofa, potted plant, bed, dining table
    Electronics:tv monitor, laptop, mouse, remote, keyboard, cell phone
    Appliances:microwave, oven, toaster, sink, refrigerator
    Other:toilet, book, clock, vase, scissors, teddy bear, hair drier, toothbrush

• [BuildingBlock] Camera Rig: added using Meta XR Building Blocks. This Game Object contains additional prefabs:
  • As child of CenterEyeAnchor:
    • DetectionUiMenuPrefab: manages and shows the UI of the sample.
• [BuildingBlock] Passthrough: Meta XR Building Blocks entity to configure and enable the Passthrough feature.
• DetectionManagerPrefab: contains the scanner logic to get the camera data and run the Inference Engine inference to update the UI elements.
• SentisInferenceManagerPrefab: contains the multi-object detection inference and UI logic.
• EnvironmentRaycastPrefab: contains the logic to use the MRUK Raycast to get the real world 3D point using the Quest Depth Data.
• PassthroughCameraAccessPrefab: creates a PassthroughCameraAccess component and manages its settings.
• ReturnToStartScene: common prefab used to go back to Samples selection.

• DetectionManager.cs: This script contains the sample logic. It gets the camera image from the PassthroughCameraAccessPrefab to send it to Inference Engine inference. Also, it manages the placement action when the user presses the A button.

• Model name: Yolov9t
• Model format: Sentis (.sentis), quantized to Uint8
• Model size: ~2.3 MB
• Classes trained: 80

• Backend: the local backend used by the Inference Engine to run the inferences
• Sentis Model: the model asset (.sentis) used for these inferences
• K_Layers Per Frame: the number of layers executed per frame.
• Label Asset: a file containing a list of objects detected by the pretrained Yolov9t model.
• UIInference: references to the SentisInferenceUiManager scripts.

• Display Image: the reference to the UI element that will contain the 2D bounding boxes.
• Box Texture: the texture used by the 2d Box.
• Box Color: the color of the bounding box.
• Font: the font type used to write the detected object information.
• Font Color: the color for the object information
• Font size: the size of the font used to show the object information.
• On Object Detected event: event launched when YOLO detects an object.

• The Inference Engine framework provides a function to convert a texture to a Tensor<float>.

• A Tensor<float> of screen coordinates (X,Y) for all detected objects.
• A Tensor<int> with the class ID for each object detected.

• IoU Threshold: YOLO IoU threshold used to identify the objects.
• Score Threshold: YOLO score threshold value used to identify the objects.

. . .
ModelQuantizer.QuantizeWeights(QuantizationType.Uint8, ref m_model);
ModelWriter.Save(FILEPATH, m_model);
. . .

• EnvironmentRayCastSampleManager script: contains the logic to perform the depth raycast using the screen position of each detected object.
• EnvironmentRaycastManager: provided by MRUK. Contains the logic to perform a raycast against the depth map generated by the Quest 3 depth sensors.

public Vector3? Raycast(Ray ray)
{
    if (EnvironmentRaycastManager.IsSupported)
    {
        if (m_raycastManager.Raycast(ray, out var hitInfo))
        {
            return hitInfo.point;
        }
        else
        {
            return null;
        }
    }
    else
    {
        Debug.LogError("EnvironmentRaycastManager is not supported");
        return null;
    }
}

• Open a console in your Windows or macOS device.
• Create a conda environment. See Installing conda⁠ for more information.

# Create conda environment
conda create -n "myenv" python=3.10
# Activate conda environment
conda activate myenv

• Install PyTorch

# Install PyTorch
pip3 install torch torchvision torchaudio

• Install the ultralytics package

# Install the ultralytics package from PyPI
pip3 install ultralytics

• Create a new python script (yolo.py) and add the following code inside

from ultralytics import YOLO
# Load a pretrained YOLO model (recommended for training)
model = YOLO("yolov9t.pt")
# Train the model using the 'coco8.yaml' dataset for 3 epochs
results = model.train(data="coco8.yaml", epochs=3)
# Evaluate the model's performance on the validation set
results = model.val()
# Export the model to ONNX format
success = model.export(format="onnx")

• Run the python script to generate the yolov9t.onnx file.

python3 yolo.py

• This script will show you the path of your onnx file when it finishes:

// ending yolo.py output sample
. . .
ONNX: starting export with onnx 1.17.0 opset 19...
ONNX: slimming with onnxslim 0.1.39...
ONNX: export success ✅ 0.8s, saved as 'runs/detect/train/weights/best.onnx' (12.2 MB)
Export complete (1.0s)
Results saved to /Users/username/yolo/runs/detect/train/weights
Predict:         yolo predict task=detect model=runs/detect/train/weights/best.onnx imgsz=640
Validate:        yolo val task=detect model=runs/detect/train/weights/best.onnx imgsz=640 data=/opt/anaconda3/envs/Executorch/lib/python3.10/site-packages/ultralytics/cfg/datasets/coco8.yaml
Visualize:       https://netron.app
. . .

1. Download the Unity Passthrough Camera API samples repository⁠ and copy this folder into a new project: Assets/Samples/PassthroughCamera.
2. Use the Unity Package Manager UI⁠ to install Inference Engine 2.2.1 into your project (com.unity.ai.inference).
3. Navigate to Meta > Tools > Project Setup Tool and fix any issues that it finds in the configuration of your project.
4. Create a new empty scene and ensure it doesn’t contain a Camera component.
5. Add an Event System to your scene.
6. Navigate to Meta > Tools > Building Blocks and add a Camera Rig.
7. To run a different AI model, follow a similar procedure to the Inference Engine sample:
   • Check the content from the Assets/PassthroughCameraApiSamples/MultiObjectDetection folder.
   • Drag the following prefabs into the project hierarchy:
     • DetectionManagerPrefab
     • SentisInferenceManagerPrefab
     • EnvironmentRaycastPrefab
     • PassthroughCameraAccessPrefab
   • Drag the DetectionUiMenuPrefab prefab to the CenterEyeAnchor. Then, drag the CenterEyeAnchor camera to the event camera for the canvas.
   • Iterate through the prefabs and ensure the fields are populated as in the sample package.
   • Build and run to verify it works.
8. The management and processing of the Inference Engine framework are handled by the following scripts:
   • DetectionManagerPrefab: see Sample detection manager prefab in the previous section. Within DetectionManager.cs, the Update() function gets the CPU texture data from the WebCamTexture object (via PassthroughCameraAccessPrefab) and kicks off the inference model by calling RunInference(captureBuffer) once the WebCamTexture object is ready. This process checks for a camera texture and waits for the current inference to complete before starting a new one. This is a standard approach that you can reuse with ML models.
   • SentisInferenceManagerPrefab: contains the critical features for running Inference Engine and can be used as a design pattern for your Inference Engine models. See YOLO control scripts in the previous section.
     • The public variables here allow you to configure the backend for the Inference Engine to run on either the GPU or the CPU. Select the Sentis model that you plan to use (Yolov9t in this case). See How to generate your own YOLO ONNX model to learn how to create these models. ONNX is the standard format accepted by Inference Engine. Finally, it lets you set the layers per frame.
     • SentisInferenceRunManager.cs controls the Inference Engine model. This includes LoadModel(), which takes the Inference Engine model, compiles the graph for the model, and spawns the worker thread that will execute the model. It contains a public function RunInference(Texture2D targetTexture), which converts the input camera image into the tensor that can be consumed by the model and kicks off the inference.
   • Modify the Inference Engine inference based on your model requirements:
     • Input data will be the same (texture to tensor). The RunInference(Texture2D targetTexture) function converts the texture to a tensor.
     • Output data will vary, so you must write your own post-processing scripts. Refer to the GetInferencesResults() function to get the output of the model using the pull request technique.
     • Use the layer by layer inference technique and pull request download function to read the output to get better performance results on Quest. See the InferenceUpdate() function to see how layer by layer inference works.
9. The UI varies depending on the model you chose. The following explains the selections used in the sample:
   • SentisInferenceUiManager.cs is the script that updates the UI based on the output of the model.
     Note: You must create your specific script to interpret the output of the model that you choose for your own application. In the case of YOLO, use DrawUIBoxes(output labelDs, imageWidth, imageHeight).
   • DetectionUiMenuPrefab prefab is under the CenterEyeAnchor and manages the display of the initial and start panels that you see as the app launches along with the countdown.

• Model architecture:
  • Use models that do not require complex architectures.
    Note: Large generative models and LLMs do not perform well on devices like Meta Quest 3.
  • Inference Engine runs on the main thread of your Unity app, so it will impact other main processes like render pipeline.
    • Use the layer by layer inference technique (split the inference in layer per frame) to prevent blocking the main thread.
• Model size:
  • The model is loaded to the CPU or GPU of your device. This can cause multiple issues when you try to use large models:
    • Long loading times
    • Lag on the main thread during the first inference
    • Reduction of the memory budget for other resources
  • Recommendations:
    • Use the smallest version of the model that you need. For example, in the MultiObject detection sample, the smallest version of the model is YOLO (8 MB), because the medium size of YOLO (146 MB) performs worse.
    • Convert the model to Inference Engine format and quantize it to Uint8 to reduce the loading times.
• GPU versus CPU:
  • Choose a suitable strategy to minimize transferring data between GPU and CPU contexts:
    • If you are using an AI model to perform graphics-related work, keep all the processing on the GPU. Select the GPUCompute backend for Inference Engine and use the procedures that operate on the GPU to send camera data directly to Inference Engine. Finally, if you don’t need to access the output of the model on the CPU, you can send it directly to your shader or material.
    • If you need to send the output to the CPU, you can:
      • Run the model on the GPU, then send results asynchronously to the CPU.
      • Run the model on the CPU.
    • Recommendations:
      • Get the output data asynchronously in any backend to prevent blocking the main thread. Inference Engine provides different techniques to accomplish this. To learn more, see Read output from a model asynchronously⁠.
• NPU backend on Inference Engine:
  • Currently, Inference Engine doesn’t use NPU or hardware acceleration to run the inferences. Inference Engine is part of Unity Engine and is designed to run on multiple platforms. For Quest devices, Inference Engine is running as a regular Android platform with no specific accelerations. Keep this in mind when selecting the model.