Skip to main content

Integrate ONE AI Models (ONNX / TensorFlow Lite)

This guide explains how to use the AI models generated by ONE AI after export. ONE AI can export models in ONNX and TensorFlow Lite formats, which can be integrated into virtually any application or platform.

Export Settings​

When exporting your model, make sure to enable "Include Pre- and Postprocessing" in the export options. This simplifies integration by embedding all necessary preprocessing (normalization, resizing) and postprocessing (result interpretation) directly into the model.

Enable Pre- and Postprocessing

Input Format​

Single Image Input​

When your model uses a single image as input, the input tensor has the following shape:

[1, 3, height, width]
DimensionDescription
1Batch size (always 1 for inference)
3RGB color channels
heightImage height in pixels (as configured during training)
widthImage width in pixels (as configured during training)

Data Type: float32 with values in range [0, 255] (when pre-processing is included)

Multiple Image Input​

For models that compare multiple images (e.g., difference detection), the input tensor shape is:

[1, 3, height, width, image_count]
DimensionDescription
1Batch size (always 1 for inference)
3RGB color channels
heightImage height in pixels
widthImage width in pixels
image_countNumber of input images

Image Order​

When working with multiple input images, the dataset typically groups images by a base name with different suffixes (e.g., img1_rgb.png and img1_depth.png). Therefore, the model expects these images to be stacked in a specific order in the input tensor [..., image_count].

The input index order depends on the suffix names:

  • Alphabetical Sorting: By default, suffixes are sorted alphabetically.
    • Example: For _rgb and _depth, _depth is assigned Index 0 and _rgb Index 1.
  • Reference Priority: If a suffix suggests a reference image (suffix is "good", "temp", "ref", or "comp", case-insensitive), it is forced to Index 0, overriding alphabetical order.
    • Note: If multiple images contain these keywords, those specific images are sorted alphabetically amongst themselves.

Output Format​

The output format depends on the task type configured during training.

Classification​

For image classification tasks, the output tensor has the shape:

[1, detected_classes, 2]
IndexDescription
0Confidence value (0.0 - 1.0)
1Class ID

Example Output:

[[0.95, 0], [0.03, 1], [0.02, 2]]

This means: Class 0 with 95% confidence, Class 1 with 3% confidence, Class 2 with 2% confidence.

Object Detection​

For object detection tasks, the output tensor has the shape:

[1, detected_objects, 6]
IndexDescription
0X center (in pixels, relative to model input size)
1Y center (in pixels, relative to model input size)
2Width (in pixels)
3Height (in pixels)
4Confidence value (0.0 - 1.0)
5Class ID

Example Output:

[[128, 96, 64, 48, 0.92, 1]]

This means: Object of Class 1 at center position (128px, 96px) with size (64px × 48px) and 92% confidence.

Important: Coordinate Reference

The pixel coordinates in the output refer to the model's input dimensions, not your original image size. If you resize or crop your image before feeding it into the model, you need to transform the coordinates back to your original image space.

Segmentation​

For semantic segmentation tasks, the output tensor has the shape:

[1, 1, height, width]

Each pixel position contains the predicted Class ID for that location.

Example: A 256x256 segmentation output would be a tensor of shape [1, 1, 256, 256] where each value represents the class at that pixel.

Integration Options​

We provide a ready-to-use C# SDK that handles all the complexity of model loading, inference, and result parsing:

C# SDK Documentation

C++ Project or Executable​

For deployment on embedded systems, servers, or when you need maximum performance, ONE AI can export a complete C++ project or precompiled executable based on TensorFlow Lite:

C++ Export Guide

C++ API Documentation

Direct Integration​

For custom integrations, you can use the standard ONNX or TensorFlow Lite runtimes:

ONNX Runtime​

TensorFlow Lite​