Why are albumentations Augmentations (Yolo / YoloV5) altering Bounding Boxes if no augmentations are being placed? - pytorch

Question

I was using the Albumentations library in order to perform some data augmentations on an object detection dataset that I intended to train a YoloV5 model on.
I have to perform the augmentations seperately and save the images locally to disk, but when I do I noticed that some of the output bounding boxes returned aren't generating properly.
I have my augmentations set up in a seperate aug.py file, shown below (augmentations purposefully removed in debugging attempts, see below) -
import albumentations as A
import cv2
PROB = 0.5
bbp = A.BboxParams(format="yolo")
horizontal_flip_transform = A.Compose([
], bbox_params = bbp)
vertical_flip_transform = A.Compose([
], bbp)
pixel_dropout_transform = A.Compose([
], bbox_params = bbp)
random_rotate = A.Compose([
], bbox_params = bbp )
#NOTE: THIS METHOD IMPLIES THAT THE IMAGE WIDTHS MUST BE AT LEAST 50 PIXELS
#Remove this aug to remove this constraint
random_crop = A.Compose([
], bbox_params = bbp)
augs = [horizontal_flip_transform, vertical_flip_transform, pixel_dropout_transform, random_rotate, random_crop]
def get_augmentations():
return augs
And the relevant parts of my implementation for performing the augmentations and saving them to disk is below:
def run_augments_on_image(img_name, bboxes, max_images_to_generate = 500):
ret = []
img = np.array(Image.open(img_name), dtype=np.uint8)
transforms = get_augmentations()
for i in range(min(len(transforms), max_images_to_generate)):
transformed = transforms[i](image=img, bboxes = bboxes)
ret.append((transformed["image"] , transformed["bboxes"]))
return ret
def run_and_save_augments_on_image_sets(batch_img_names, bboxes_urls, max_images_to_generate, dataset_dir, trainval):
num_images = 0
for i in range(len(batch_img_names)):
bboxes = []
with open(os.path.join(dataset_dir, trainval, 'labels', bboxes_urls[i]), 'r') as f:
for row in f:
x = row.strip().split(' ')
x.append(row[0])
x.pop(0)
x[0] = float(x[0])
x[1] = float(x[1])
x[2] = float(x[2])
x[3] = float(x[3])
bboxes.append(x)
trans = run_augments_on_image(os.path.join(dataset_dir, trainval, 'images', batch_img_names[i]), bboxes)
img_index = len(os.listdir(os.path.join(dataset_dir, 'train' , 'images'))) + len(os.listdir(os.path.join(dataset_dir, 'valid', 'images'))) + 1
for j in range(len(trans)):
img_trans, bboxes_trans = trans[j]
p = Image.fromarray(img_trans).save(os.path.join(dataset_dir, trainval, 'images' , f'image-{img_index}.{batch_img_names[j].split(".")[-1]}'))
with open(os.path.join(dataset_dir, trainval, 'labels', f'image-{img_index}.txt'), 'w') as f:
for boxs in bboxes_trans:
print(f'{boxs[-1]} {boxs[0]} {boxs[1]} {boxs[2]} {boxs[3]}', file=f)
num_images += 1
img_index += 1
if num_images >= max_images_to_generate:
break
if num_images >= max_images_to_generate:
break
For testing purposes (some of the bounding boxes were off), I removed all the actual augmentations, expecting the input image label (one augmented image example shown below) to be equal to augmented label since there were no augmentations. But, as you can see, the two labels are different.
img-original.txt
0 0.5662285714285714 0.2740066225165563 0.5297714285714286 0.4837913907284769
img-augmented.txt
0 0.51488 0.47173333333333334 0.6405099999999999 0.6527333333333334
(The labels above are in normalized xywh YOLO format)
Why is albumentations altering the labels? None of the augmentations in augs.py contain anything.