Image rotation action, usable in chains

[beniz]

Image rotation action, typically useful prior to an OCR model that requires an element with text is correctly rotated.

The rotate action is to be used as follows:

• within a chain:

{                                                                                                     
                 "action": {                                                                                          
                     "type": "rotate"                                                                                 
                 }                                                                                                    
}

• action takes the output label of a classifier model that must be one of the following strings: ["0","90","180","270"] as the rotation angle in degrees. The output of the action is then the image (or object crop depending on the chain) rotated by the angle predicted by the model.

• action has parameter orientation with values as "relative" (default) or "absolute":

  • "relative": the output orientation label yields the image/crop orientation, e.g. 90, and must be corrected relative to this value in order to get rotate to 0, e.g. with -90 = 270 rotation.
  • "absolute": the output orientation is the command to the rotation to be applied.