Inference output

Inference output examples

Let's first start with some examples of output which may be easier to understand than the exact specifications below.

A typical output for a classification algorithm will look like this:

JSON

{
    "outputs": [{
        "labels": {
            "predicted": [{
                "label_name": "sunglasses",
                "label_id": 9,
                "score": 0.801353174,
                "threshold": 0.347
            },
            {
                "label_name": "glasses",
                "label_id": 8,
                "score": 0.175235228,
                "threshold": 0.139
            }],
            "discarded": []
        }
    }]
}

Prediction are displayed as soon as the prediction score is above the indicated threshold. The predictions are sorted by decreasing confidence, so you may only be interested in the first prediction outputs[0]['labels']['predicted'][0]['label_name'] for a classification problem. For a multi-label classification problem (tagging), all the predicted labels are independent and may be of interest.

The output of a detection algorithm is almost the same, see just below. Please note the presence of the roi field which indicate the location ("roi" stand for Region Of Interest) of the object in the image. The (xmin, ymin) coordinate of the box corresponds to its top-left corner.

JSON

{
    "outputs": [{
        "labels": {
            "predicted": [{
                "label_name": "sunglasses",
                "label_id": 9,
                "roi": {
                    "region_id": 1,
                    "bbox": {
                        "xmin": 0.312604159,
                        "ymin": 0.366485775,
                        "ymax": 0.5318923,
                        "xmax": 0.666821837
                    }
                },
                "score": 0.990304172,
                "threshold": 0.347
            }],
            "discarded": []
        }
    }]
}

Inference output specification

The data returned by an inference request will contain the list of inference outputs:

Attribute	Type	Description
outputs	array(object)	An array of inference output objects. The i-th element of the array corresponds to the result of the i-th element of the specification outputs field and version post_processings field.

This object is directly related to the output object of the specification: they both have the same unique field. It stores the recognition inference output.

Attribute	Type	Description
labels	object	An output of type labels.

Inference labels output

Attribute	Type	Description
predicted	array(object)	An array of prediction object. This field stores the list of recognition hypotheses whose score is above the recognition threshold.
discarded	array(object)	An array of prediction object. If you passed show_discared=true in the inference request, this field will store the list of recognition hypotheses whose score did not reached the recognition threshold.

Prediction

Stores information related to an object hypothesis.

Attribute	Type	Description
label_id	int	The recognized label ID from the specification's label object.
label_name	string	The recognized label name from the specification's label object.
score	float	The recognition score.
threshold	float	The recognition threshold that was defined by the recognition version post-processing.
roi	object	(optional) If the roi field of the corresponding labels output object is not "NONE", this field will store a ROI object.

ROI Object

ROI stands for Region Of Interest and describes the position of an object.

Attribute	Type	Description
region_id	int	The region ID. It might not be unique among all the returned prediction objects. It can be used in conjunction with show_discarded=true to group the predicted and discarded fields by region_id to identify alternate labels for a given region in case of ambiguity.
bbox	object	(optional) Present if and only if the region type is "BBOX".