resize images with bboxes for training
#3
by
abhishek
HF staff
- opened
Using the package chitra, you can easily resize images and bounding boxes at the same time:
Example:
from chitra.image import Chitra
import pandas as pd
import matplotlib.pyplot as plt
train_df = pd.read_csv("train.csv")
id = 73
train_df = train_df[train_df["id"] == f"{id}.png"].reset_index(drop=True)
bboxes = [k for k in train_df[['xmin', 'ymin', 'xmax', 'ymax']].values]
image = Chitra(f"train/{id}.png", bboxes=bboxes, labels=[0 for _ in range(len(bboxes))])
image.resize_image_with_bbox((2048, 2048))
plt.imshow(image.draw_boxes())
Please note that for your submission file, you must submit bboxes for the original image sizes in the test set.
Alternatively, we can use the albumentations library and provide it the bounding box parameters to resize images along with their bounding boxes.
Example:
import albumentations as A
from albumentations.pytorch import ToTensorV2
def get_transforms(cfg, data):
"""
Function to obtain the Training and Validation Transforms
Args:
cfg (dict): Configuration File
data (str): "train" for Training data,
"valid " for Validation data
Returns:
Augmentations : Transforms to be applied
"""
# Train Augmentations
if data == "train":
return A.Compose(
[
A.Resize(cfg['MODEL']['IMAGE_SIZE'], cfg['MODEL']['IMAGE_SIZE']),
A.HorizontalFlip(p=0.5),
A.RandomBrightnessContrast(p=0.5),
ToTensorV2()
],
p=1.,
bbox_params={'format': 'pascal_voc', 'label_fields': ['labels']}
)
# Validation Augmentations
if data == "valid":
return A.Compose(
[
A.Resize(cfg['MODEL']['IMAGE_SIZE'], cfg['MODEL']['IMAGE_SIZE']),
ToTensorV2()
],
p=1.,
bbox_params={'format': 'pascal_voc', 'label_fields': ['labels']}
)
def get_test_transform():
return A.Compose([
ToTensorV2(p=1.0)
],
)
@abhishek is there any difference between using this method and using chitra which you may have observed?
@ishandutta
I am using your method for resizing the images.
And then I am using this for applying transforms to the images and bbox
if self.transform is not None:
transformed = self.transform(image=img, bboxes=target['boxes'], class_labels=target['labels'])
image = transformed['image']
target['boxes'] = transformed['bboxes']
target['labels'] = transformed['labels']
But I am getting this error
Can you share how you are applying the transform