Image Letterboxing (or Pillarboxing)

img_letterbox.py

import cv2
import numpy as np


__all__ = ['letterbox']


def letterbox(img_arr, target_size, boxes=None):
    """Image Letterboxing (or Pillarboxing)
    
    Args:
        img_arr (np.ndarray dtype=uint8)
        target_size (tuple): (height: int, width: int)
        boxes (np.ndarray dtype=np.int32): shape (n, 4) --> [x_min, y_min, x_max, y_max]
    
    """
    img_arr = img_arr.copy()
    if boxes:
        boxes = boxes.copy()
    
    input_height, input_width = img_arr.shape[:2]
    target_height, target_width = target_size

    scale = min(target_width / input_width, target_height / input_height)
    resized_width, resized_height  = int(scale * input_width), int(scale * input_height)
    img_resized = cv2.resize(img_arr, dsize=(resized_width, resized_height))

    half_pad_width, half_pad_height = (target_width - resized_width) // 2, (target_height - resized_height) // 2
    img_padded = np.zeros(shape=(target_height, target_width, 3), dtype=np.uint8)
    img_padded[half_pad_height:half_pad_height+resized_height, half_pad_width:half_pad_width+resized_width, :] = img_resized

    if not boxes:
        return img_padded
    else:
        boxes[:, [0, 2]] = boxes[:, [0, 2]] * scale + half_pad_width
        boxes[:, [1, 3]] = boxes[:, [1, 3]] * scale + half_pad_height
        return img_padded, boxes