Specific export config details as per image recognition type#

User must provide an image recognition config before starting the data export. Depending upon the type of image recognition whether it is a data export for object detection data or for classification data, user can select the following options.

# sample 'classification' type config
            "img_recognition_configs": [
                {
                  "type": "classification",
                  "export_format": "imagenet",                    
                  "classification_threshold": {"anomalous":0.5},
                  "filter": {
                      "annotation_type": ["auto", "human"],                   
                      "start_date": "2023-06-06T00:00:00Z",
                      "end_date": "2024-12-30T13:00:00Z"
                      }
                }
            ]
# sample 'object_detection' type config
            "img_recognition_configs": [  
              {
                  "type": "object_detection",
                  "export_format": "COCO",
                  "filter": {
                      "label_score": {"Person": 0.45, "vehicle": 0.6},
                      "annotation_type": ["auto", "human"],
                      "start_date": "2023-06-06T00:00:00Z",
                      "end_date": "2024-12-30T13:00:00Z"
                      }
              }
      ]

Common#

  • type: Mandatory. Refers to type of image recognition config. Can be either classification or object_detection. Both type of configs together is not supported at the moment due to database schema challenges.

  • export_format: Optional. Defaults as per the type selection. Refers to data export format. For classification, currently only imagenet is supported. Similarly, for object detection COCO format is supported.

  • filter: Optional. sub config for data filter. Data in DataStore satisfying these filter combination (AND of all the options provided) will only be exported.

    • annotation_type: list of type(s) of annotation of the metadata. can be either auto,human or both.

    • start_date: UTC timestamp of start date of ingestion

    • end_date: UTC timestamp of end date of ingestion. Must if start_date is also provided

    • Note: Ensure that both start_date and end_date are within the retention policy of DataStore.

    All filter values are optional. If omitted, they just won’t be the part of where clause in search query thus widening the search criteria.

Recognition type specific#

  • CLASSIFICATION

    • classification_threshold: Optional. Defaults to 0.5. A key-value pair referring to class label that will set the threshold >= which data is exported. For example, a metadata has a label named anomalous score of 0.45, and in the config, classification_threshold is set to 0.5, the export for this particular metadata will be ignored

  • OBJECT_DETECTION

    • filter.label_score: Optional. A key-value pair referring to bounding box labels and their score, >= which the frame metadata export will be exported. Any one bbox if found violating the label-score, the export for entire frame is dropped.

Unsupported features#

Both object detection(OD) and classification(CLF) data export are supported with following exclusions.

  • Unsupported features for CLF

    • device type azure is not supported for type is classification.

Important#

Also, ensure that the labels provided in classification_threshold or in label_score in annotation.categories list in Intel® Edge Data Collection config.(e.g. anomalous in the given example config.)