A virtual column is a column whose values are computed during query execution using a chosen function applied over real columns in the catalog table.
The video below shows creating a view with virtual columns. The catalog table has a bounding box as a JSON string with 4 corner co-ordinates in the form of {"x_left":<>, "x_right":<>, "y_top":<>, "y_bottom":<>}.
- The bounding box area is defined on a column of type bounding box.
- The bounding box type required for the above definition is defined on the column that contains the bounding box coordinates as a JSON string.
Once the view is saved, the virtual column acts like any real column and is available as a column that can be used in query filter conditions like any other real column. For example, the gt_area column created in the above video can be used in query conditions like (gt_area > 0.5) to filter out rows with a bounding box that occupies more than 50% of the image.
Virtual column types for 2D object detection
| Type name | Output | Input |
AKD_2DBoundingBox
AKD_2DBoundinBoxAbs
AKD_2DBoundingBoxCOCOJSON
AKD_2DBoundingBoxJSON
| A 2D bounding box defined by return type boundingbox2d | - AKD_2DBoundingBox - Takes four individual coordinate columns with coordinate values normalized between 0 and 1.
- AKD_2DBoundinBoxAbs - Takes four individual coordinate columns with absolute pixel values and an image height and width column.
- AKD_2DBoundingBoxCOCOJSON - Takes a single column that holds bounding box in COCO bounding box format([x_left, y_top, box_width, box_height]) and image height and width columns.
- AKD_2DBoundingBoxJSON - Takes a single column with a JSON string formatted bounding box coordinates normalized between 0 and 1. The JSON is in the {x_left:<>, x_right:<>, y_top:<>, y_bottom:<>} format.
|
| AKD_2DBoundingBoxGT | A 2D object detection ground truth with class and bounding box. | - 2D Bounding box of type boundingbox2d specified using any supported input types described above.
- A class label of type string or a mapped class type classlabelgt.
|
| AKD_2DBoundingBoxPred | A 2D object detection prediction with class, bounding box, and prediction confidence. | - Input set 1
- 2D Bounding box of type boundingbox2d specified using any supported input types described above.
- A class label of type string or a mapped class type classlabelpred.
- Prediction confidence of type float or double.
- Input set 2
- 2D Bounding box of type boundingbox2d specified using any supported input types described above.
- A column that is defined using classlabelwithconfidence type that combines class label and prediction confidence.
|
| AKD_2DBoundingBoxArea | Area of a 2D bounding box | - 2D Bounding box of type boundingbox2d specified using any supported input types described above.
|
Virtual column types for segmentation
| Type name | Output | Input |
| AKD_SegmentJson | An instance segment represented by the type 'segment'. | - A column with JSON that represents the segment in polygon format as [x1, y1, x2, y2,...] with (x1, y1), (x2, y2) etc being the vertex coordinates for polygon making up the segment.
- mask_width and mask_height columns that represent the width and height of the image used to define the segment coordinates.
|
| AKD_SegmentGt | An instance segment with a ground truth. | - A column of type 'segment'. Currently, AKD_SegmentJson above should be used to define a segment.
- A column with a ground truth label.
|
| AKD_SegmentPred | An instance segment with a prediction. | - Input set 1
- A column of type 'segment'. Currently, AKD_SegmentJson above should be used to define a segment.
- A class label of type string or a mapped class type classlabelpred.
- Prediction confidence of type float or double.
- Input set 2
- A column of type 'segment'. Currently, AKD_SegmentJson above should be used to define a segment.
- A column that is defined using classlabelwithconfidence type that combines class label and prediction confidence.
|