Interpreting results

run returns one ImageInferenceResult per input image. Each holds a flat list of InferenceResult entries under results. In C#, these records are exposed through the disposable, list-like ImageInferenceResults owner returned by Run.

InferenceResult

Each InferenceResult carries the output of one network stage. The populated field depends on network_type:

Field	Populated for
bounding_box	object detection, instance segmentation
segmentation	segmentation, instance segmentation, anomaly detection
scalar	classification, anomaly detection
text	optical character recognition
barcode	barcode reading
sub_results	nested results from downstream stages

Additional fields: network_type (the producing stage), index (position within its sibling list), and batch_index (the source image). Coordinates on bounding_box, segmentation and barcode are absolute pixel values. C# exposes the same fields as PascalCase record properties.

Object definition

Python

class InferenceResult:
    network_type: str
    batch_index: int
    index: int
    sub_results: list[InferenceResult]
    scalar: Scalar | None
    bounding_box: BoundingBox | None
    segmentation: Segmentation | None
    text: str | None
    barcode: Barcode | None

C#

public sealed class ImageInferenceResults :
    IReadOnlyList<ImageInferenceResult>,
    IDisposable
{
    public IReadOnlyList<ImageInferenceResult> Results { get; }
    // Count, indexer, enumeration, and Dispose are also available.
}

public sealed record ImageInferenceResult(
    IReadOnlyList<InferenceResult> Results);

public sealed record InferenceResult(
    NetworkType NetworkType,
    ulong BatchIndex,
    ulong Index,
    IReadOnlyList<InferenceResult> SubResults,
    BoundingBox? BoundingBox,
    SegmentationResult? Segmentation,
    Scalar? Scalar,
    string? Text,
    Barcode? Barcode);

ImageInferenceResults retains the native result tree needed for ToImagesWithAnnotations and must be disposed. Its ImageInferenceResult and InferenceResult records are managed copies and remain valid after the owner is disposed.

C / C++

typedef struct DenkflowInferenceResult {
    DenkflowNetworkType network_type;
    size_t index;
    struct DenkflowInferenceResult *sub_results;
    size_t sub_results_length;
    DenkflowBoundingBox *bounding_box;
    DenkflowSegmentation *segmentation;
    DenkflowScalar *scalar;
    char *text;
    DenkflowBarcode *barcode;
} DenkflowInferenceResult;

Walking the results

Python

for image_result in results:
    for r in image_result.results:
        if r.bounding_box:
            print(r.bounding_box.class_label.name, r.bounding_box.confidence)
        if r.text is not None:
            print(r.text)
        for sub in r.sub_results:
            ...  # nested stage output

C#

foreach (ImageInferenceResult imageResult in results)
{
    foreach (InferenceResult result in imageResult.Results)
    {
        if (result.BoundingBox is BoundingBox box)
        {
            Console.WriteLine($"{box.ClassLabel.Name}: {box.Confidence:F2}");
        }

        if (result.Text is string text)
        {
            Console.WriteLine(text);
        }

        foreach (InferenceResult subResult in result.SubResults)
        {
            // Handle output from a nested stage.
        }
    }
}

C / C++

for (size_t b = 0; b < inference_results->image_results_length; ++b) {
	DenkflowImageInferenceResult* image_result = &inference_results->image_results[b];
	for (size_t i = 0; i < image_result->results_length; ++i) {
		DenkflowInferenceResult* r = &image_result->results[i];
		if (r->bounding_box != NULL) {
			printf("%s: %f\n", r->bounding_box->class_label.name, r->bounding_box->confidence);
		}
	}
}

In C, the result tree is released once with denkflow_inference_results_free.

Nested pipelines (for example detection followed by per-object classification) expose the downstream output under sub_results.