Stitch ocr detections workflow block

[reiffd7]

Description

I created a transformation workflow block called "Stitch OCR Detections". At its core, this block combines detection class names into a text string based on where the detections are located spatially. It will be useful for object detection OCR models where class names are characters. It allows for user input to dictate how the results will look based on language of the text and size of the image.

This transformation takes OCR detection results and reconstructs the original text by:

1. Grouping text detections into rows based on their vertical (y) positions
2. Sorting detections within each row by horizontal (x) position
3. Concatenating the detected text in reading order

The block supports two configurable parameters:

Reading Direction (dropdown)

• "left_to_right": Standard left-to-right reading (e.g., English)
• "right_to_left": Right-to-left reading (e.g., Arabic)
• "vertical_top_to_bottom": Vertical reading from top to bottom
• "vertical_bottom_to_top": Vertical reading from bottom to top

Tolerance (integer)

Controls how close detections need to be vertically (in pixels) to be considered part of the same line of text. A higher tolerance will group detections that are further apart vertically.

This block is particularly useful for:

• Converting individual character/word detections into readable text
• Reconstructing multi-line text from OCR results
• Maintaining proper reading order of detected text elements
• Supporting different writing systems and text orientations

Type of change

• New feature (non-breaking change which adds functionality)

Testing

The changes have been tested through:

1. Unit Tests

   • Tests for all reading directions
   • Edge cases (empty detections, single characters)
   • Tolerance grouping behavior
   • Multi-line text handling

2. Integration Testing

   • Tested on local inference server
   • Created workflows with various text orientations:
     • Horizontal text
     • Vertical text
     • Multi-line text
     • Different languages/writing systems

[CLAassistant]

All committers have signed the CLA.