Working with CVAT Bounding Box Annotations in Torchvision

Learn how to work with CVAT bounding box annotations in torchvision for object detection tasks.

Christian Mills


January 21, 2024

This post is part of the following series:


Welcome to this hands-on guide for working with bounding box annotations created with the CVAT annotation tool in torchvision. Bounding box annotations specify rectangular frames around objects in images to identify and locate them for training object detection models.

The tutorial walks through setting up a Python environment, loading the raw annotations into a Pandas DataFrame, annotating and augmenting images using torchvision’s Transforms V2 API, and creating a custom Dataset class to feed samples to a model.

This guide is suitable for beginners and experienced practitioners, providing the code, explanations, and resources needed to understand and implement each step. By the end, you will have a solid foundation for working with bounding box annotations made with CVAT for object detection tasks.

Getting Started with the Code

The tutorial code is available as a Jupyter Notebook, which you can run locally or in a cloud-based environment like Google Colab. I have dedicated tutorials for those new to these platforms or who need guidance setting up:

Tutorial Code
Jupyter Notebook: GitHub Repository Open In Colab

Setting Up Your Python Environment

Before diving into the code, we’ll cover the steps to create a local Python environment and install the necessary dependencies.

Creating a Python Environment

First, we’ll create a Python environment using Conda/Mamba. Open a terminal with Conda/Mamba installed and run the following commands:

# Create a new Python 3.10 environment
conda create --name pytorch-env python=3.10 -y
# Activate the environment
conda activate pytorch-env
# Create a new Python 3.10 environment
mamba create --name pytorch-env python=3.10 -y
# Activate the environment
mamba activate pytorch-env

Installing PyTorch

Next, we’ll install PyTorch. Run the appropriate command for your hardware and operating system.

# Install PyTorch with CUDA
pip install torch torchvision torchaudio --index-url
# MPS (Metal Performance Shaders) acceleration is available on MacOS 12.3+
pip install torch torchvision torchaudio
# Install PyTorch for CPU only
pip install torch torchvision torchaudio --index-url
# Install PyTorch for CPU only
pip install torch torchvision torchaudio

Installing Additional Libraries

We also need to install some additional libraries for our project.

Package Description
jupyter An open-source web application that allows you to create and share documents that contain live code, equations, visualizations, and narrative text. (link)
matplotlib This package provides a comprehensive collection of visualization tools to create high-quality plots, charts, and graphs for data exploration and presentation. (link)
pandas This package provides fast, powerful, and flexible data analysis and manipulation tools. (link)
pillow The Python Imaging Library adds image processing capabilities. (link)
tqdm A Python library that provides fast, extensible progress bars for loops and other iterable objects in Python. (link)
distinctipy A lightweight python package providing functions to generate colours that are visually distinct from one another. (link)

Run the following commands to install these additional libraries:

# Install additional dependencies
pip install distinctipy jupyter matplotlib pandas pillow tqdm

Installing Utility Packages

We will also install some utility packages I made, which provide shortcuts for routine tasks.

Package Description
cjm_pil_utils Some PIL utility functions I frequently use. (link)
cjm_psl_utils Some utility functions using the Python Standard Library. (link)
cjm_pytorch_utils Some utility functions for working with PyTorch. (link)
cjm_torchvision_tfms Some custom Torchvision tranforms. (link)

Run the following commands to install the utility packages:

# Install additional utility packages
pip install cjm_pil_utils cjm_psl_utils cjm_pytorch_utils cjm_torchvision_tfms

With our environment set up, we can open our Jupyter Notebook and dive into the code.

Importing the Required Dependencies

First, we will import the necessary Python packages into our Jupyter Notebook.

# Import Python Standard Library dependencies
from functools import partial
from pathlib import Path
import xml.etree.ElementTree as ET

# Import utility functions
from cjm_pil_utils.core import get_img_files
from cjm_psl_utils.core import download_file, file_extract
from cjm_pytorch_utils.core import tensor_to_pil
from cjm_torchvision_tfms.core import ResizeMax, PadSquare, CustomRandomIoUCrop

# Import the distinctipy module
from distinctipy import distinctipy

# Import matplotlib for creating plots
import matplotlib.pyplot as plt

# Import numpy
import numpy as np

# Import the pandas package
import pandas as pd

# Do not truncate the contents of cells and display all rows and columns
pd.set_option('max_colwidth', None, 'display.max_rows', None, 'display.max_columns', None)

# Import PIL for image manipulation
from PIL import Image

# Import PyTorch dependencies
import torch
from import Dataset, DataLoader

# Import torchvision dependencies
import torchvision
from torchvision.tv_tensors import BoundingBoxes
from torchvision.utils import draw_bounding_boxes
import torchvision.transforms.v2  as transforms

# Import tqdm for progress bar
from import tqdm

Torchvision provides dedicated torch.Tensor subclasses for different annotation types called TVTensors. Torchvision’s V2 transforms use these subclasses to update the annotations based on the applied image augmentations. The TVTensor class for bounding box annotations is called BoundingBoxes. Torchvision also includes a draw_bounding_boxes function to annotate images.

Loading and Exploring the Dataset

After importing the dependencies, we can start working with our data. I annotated a toy dataset with bounding boxes for this tutorial using images from the free stock photo site Pexels. The dataset is available on HuggingFace Hub at the link below:

Setting the Directory Paths

We first need to specify a place to store our dataset and a location to download the zip file containing it. The following code creates the folders in the current directory (./). Update the path if that is not suitable for you.

# Define path to store datasets
dataset_dir = Path("./Datasets/")
# Create the dataset directory if it does not exist
dataset_dir.mkdir(parents=True, exist_ok=True)

# Define path to store archive files
archive_dir = dataset_dir/'../Archive'
# Create the archive directory if it does not exist
archive_dir.mkdir(parents=True, exist_ok=True)

# Creating a Series with the paths and converting it to a DataFrame for display
    "Dataset Directory:": dataset_dir, 
    "Archive Directory:": archive_dir
Dataset Directory: Datasets
Archive Directory: Datasets/../Archive

Setting the Dataset Path

Next, we construct the name for the Hugging Face Hub dataset and set where to download and extract the dataset.

# Set the name of the dataset
dataset_name = 'cvat-bounding-box-toy-dataset'

# Construct the HuggingFace Hub dataset name by combining the username and dataset name
hf_dataset = f'cj-mills/{dataset_name}'

# Create the path to the zip file that contains the dataset
archive_path = Path(f'{archive_dir}/{dataset_name}.zip')

# Create the path to the directory where the dataset will be extracted
dataset_path = Path(f'{dataset_dir}/{dataset_name}')

# Creating a Series with the dataset name and paths and converting it to a DataFrame for display
    "HuggingFace Dataset:": hf_dataset, 
    "Archive Path:": archive_path, 
    "Dataset Path:": dataset_path
HuggingFace Dataset: cj-mills/cvat-bounding-box-toy-dataset
Archive Path: Datasets/../Archive/
Dataset Path: Datasets/cvat-bounding-box-toy-dataset

Downloading the Dataset

We can now download the archive file and extract the dataset using the download_file and file_extract functions from the cjm_psl_utils package. We can delete the archive afterward to save space.

# Construct the HuggingFace Hub dataset URL
dataset_url = f"{hf_dataset}/resolve/main/{dataset_name}.zip"
print(f"HuggingFace Dataset URL: {dataset_url}")

# Set whether to delete the archive file after extracting the dataset
delete_archive = True

# Download the dataset if not present
if dataset_path.is_dir():
    print("Dataset folder already exists")
    print("Downloading dataset...")
    download_file(dataset_url, archive_dir)    
    print("Extracting dataset...")
    file_extract(fname=archive_path, dest=dataset_dir)
    # Delete the archive if specified
    if delete_archive: archive_path.unlink()

Getting the Images and Annotations

The dataset has a folder containing the sample images and an XML file containing the annotations.

 # Assuming the images are stored in a subfolder named 'default'
img_dir = dataset_path/'images/default'

# Assuming annotation file is in XML format and located in any subdirectory of the dataset
annotation_file_path = dataset_path/'annotations.xml'

# Creating a Series with the paths and converting it to a DataFrame for display
    "Image Folder": img_dir, 
    "Annotation File": annotation_file_path}).to_frame().style.hide(axis='columns')
Image Folder Datasets/cvat-bounding-box-toy-dataset/images/default
Annotation File Datasets/cvat-bounding-box-toy-dataset/annotations.xml

Get Image File Paths

Each image file has a unique name that we can use to locate the corresponding annotation data. We can make a dictionary that maps image names to file paths. The dictionary will allow us to retrieve the file path for a given image more efficiently.

# Get all image files in the 'img_dir' directory
img_dict = {
    file.stem : file # Create a dictionary that maps file names to file paths
    for file in get_img_files(img_dir) # Get a list of image files in the image directory

# Print the number of image files
print(f"Number of Images: {len(img_dict)}")

# Display the first five entries from the dictionary using a Pandas DataFrame
pd.DataFrame.from_dict(img_dict, orient='index').head()
Number of Images: 28
258421 Datasets/cvat-bounding-box-toy-dataset/images/default/258421.jpg
3075367 Datasets/cvat-bounding-box-toy-dataset/images/default/3075367.jpg
3076319 Datasets/cvat-bounding-box-toy-dataset/images/default/3076319.jpg
3145551 Datasets/cvat-bounding-box-toy-dataset/images/default/3145551.jpg
3176048 Datasets/cvat-bounding-box-toy-dataset/images/default/3176048.jpg

Get Image Annotations

Next, we read the content of the XML annotation file into a Pandas DataFrame so we can easily query the annotations.

Define a function to parse the CVAT XML annotations

The following helper function parses the raw XML content into a Pandas DataFrame.

def parse_cvat_bbox_xml(xml_content):
    Parse the given XML content of a CVAT bounding box annotation file and
    convert it into a pandas DataFrame.

    This function processes the XML content to extract information about each
    image and its associated bounding boxes. The data is then structured into
    a DataFrame for easier manipulation and analysis.

    xml_content (str): A string containing the XML content from a CVAT bounding box
                       annotation file.

    pandas.DataFrame: A DataFrame where each row represents an image and contains
                      the following columns:
                      - Image ID (int): The unique identifier of the image.
                      - Image Name (str): The name of the image.
                      - Width (int): The width of the image in pixels.
                      - Height (int): The height of the image in pixels.
                      - Boxes (list): A list of dictionaries, where each dictionary
                                      represents a bounding box with keys 'Label',
                                      'xtl', 'ytl', 'xbr', and 'ybr' indicating the
                                      label and coordinates of the box.

    # Parse the XML content
    root = ET.fromstring(xml_content)
    data = {}

    # Iterate through each image element in the XML
    for image in root.findall('image'):
        # Extract image attributes
        image_id = image.get('id')
        image_name = image.get('name')
        width = image.get('width')
        height = image.get('height')

        # Initialize a dictionary to store image data
        image_data = {
            'Image ID': int(image_id),
            'Image Name': image_name,
            'Width': int(width),
            'Height': int(height),
            'Boxes': []

        # Iterate through each bounding box element within the image
        for box in image.findall('box'):
            # Extract box attributes
            label = box.get('label')
            xtl = float(box.get('xtl'))
            ytl = float(box.get('ytl'))
            xbr = float(box.get('xbr'))
            ybr = float(box.get('ybr'))

            # Construct a dictionary for the box data
            box_data = {
                'Label': label,
                'xtl': xtl,
                'ytl': ytl,
                'xbr': xbr,
                'ybr': ybr

            # Append the box data to the image's 'Boxes' list

        # Map the image data to its ID in the data dictionary
        data[image_id] = image_data

    # Convert the data dictionary to a DataFrame and return
    return pd.DataFrame.from_dict(data, orient='index')

Load CVAT XML annotations into a DataFrame

After parsing the XML content, we will change the index for the annotation_df DataFrame to match the keys in the img_dict dictionary, allowing us to retrieve both the image paths and annotation data using the same index key.

# Read the XML file
with open(annotation_file_path, 'r', encoding='utf-8') as file:
    xml_content =

# Parse the XML content
annotation_df = parse_cvat_bbox_xml(xml_content)

# Add a new column 'Image ID' by extracting it from 'Image Name'
# This assumes that the 'Image ID' is the part of the 'Image Name' before the first period
annotation_df['Image ID'] = annotation_df['Image Name'].apply(lambda x: x.split('.')[0])

# Set the new 'Image ID' column as the index of the DataFrame
annotation_df = annotation_df.set_index('Image ID')

# Display the first few rows of the DataFrame
Image Name Width Height Boxes
Image ID
258421 258421.jpg 768 1152 [{‘Label’: ‘person’, ‘xtl’: 386.08, ‘ytl’: 443.94, ‘xbr’: 460.82, ‘ybr’: 777.16}, {‘Label’: ‘person’, ‘xtl’: 340.25, ‘ytl’: 466.94, ‘xbr’: 418.99, ‘ybr’: 777.34}]
3075367 3075367.jpg 1344 768 [{‘Label’: ‘person’, ‘xtl’: 413.32, ‘ytl’: 41.22, ‘xbr’: 919.81, ‘ybr’: 763.17}]
3076319 3076319.jpg 768 1120 [{‘Label’: ‘person’, ‘xtl’: 335.31, ‘ytl’: 151.75, ‘xbr’: 711.22, ‘ybr’: 1117.49}, {‘Label’: ‘person’, ‘xtl’: 8.11, ‘ytl’: 131.88, ‘xbr’: 404.2, ‘ybr’: 1119.0}]
3145551 3145551.jpg 1184 768 [{‘Label’: ‘person’, ‘xtl’: 642.0, ‘ytl’: 289.85, ‘xbr’: 669.66, ‘ybr’: 398.89}, {‘Label’: ‘person’, ‘xtl’: 658.63, ‘ytl’: 281.25, ‘xbr’: 687.09, ‘ybr’: 398.61}]
3176048 3176048.jpg 1152 768 [{‘Label’: ‘person’, ‘xtl’: 518.23, ‘ytl’: 338.97, ‘xbr’: 594.63, ‘ybr’: 466.08}, {‘Label’: ‘person’, ‘xtl’: 683.42, ‘ytl’: 356.48, ‘xbr’: 638.86, ‘ybr’: 437.82}]

The source XML content corresponding to the first row in the DataFrame is available below:

<?xml version="1.0" encoding="utf-8"?>
  <image id="0" name="258421.jpg" subset="default" task_id="7" width="768" height="1152">
    <box label="person" source="file" occluded="0" xtl="386.08" ytl="443.94" xbr="460.82" ybr="777.16" z_order="0">
    <box label="person" source="file" occluded="0" xtl="340.25" ytl="466.94" xbr="418.99" ybr="777.34" z_order="0">

The the xtl, ytl, xbr, and ybr values indicate the bounding box annotations are in [Top-Left X, Top-Left Y, Bottom-Right X, Bottom-Right Y] format.

With the annotations loaded, we can start inspecting our dataset.

Inspecting the Class Distribution

First, we get the names of all the classes in our dataset and inspect the distribution of samples among these classes. This step won’t yield any insights for the toy dataset but is worth doing for real-world projects. A balanced dataset (where each class has approximately the same number of instances) is ideal for training a machine-learning model.

Get image classes

# Explode the 'boxes_df' column in the annotation_df dataframe
# Convert the resulting series to a dataframe and rename the 'boxes_df' column to 'boxes_df'
# Apply the pandas Series function to the 'boxes_df' column of the dataframe
boxes_df = annotation_df['Boxes'].explode().to_frame().Boxes.apply(pd.Series)

# Get a list of unique labels in the 'annotation_df' DataFrame
class_names = boxes_df['Label'].unique().tolist()

# Display labels using a Pandas DataFrame
0 person

Visualize the class distribution

# Get the number of samples for each object class
class_counts = boxes_df['Label'].value_counts()

# Plot the distribution
class_counts.plot(kind='bar', figsize=(12, 5))
plt.title('Class distribution')
plt.xticks(range(len(class_counts.index)), class_counts.index, rotation=75)  # Set the x-axis tick labels

Visualizing Image Annotations

In this section, we will annotate a single image with its bounding boxes using torchvision’s BoundingBoxes class and draw_bounding_boxes function.

Generate a color map

While not required, assigning a unique color to bounding boxes for each object class enhances visual distinction, allowing for easier identification of different objects in the scene. We can use the distinctipy package to generate a visually distinct colormap.

# Generate a list of colors with a length equal to the number of labels
colors = distinctipy.get_colors(len(class_names))

# Make a copy of the color map in integer format
int_colors = [tuple(int(c*255) for c in color) for color in colors]

# Generate a color swatch to visualize the color map

Download a font file

The draw_bounding_boxes function included with torchvision uses a pretty small font size. We can increase the font size if we use a custom font. Font files are available on sites like Google Fonts, or we can use one included with the operating system.

# Set the name of the font file
font_file = 'KFOlCnqEu92Fr1MmEU9vAw.ttf'

# Download the font file
download_file(f"{font_file}", "./")

Define the bounding box annotation function

We can make a partial function using draw_bounding_boxes since we’ll use the same box thickness and font each time we visualize bounding boxes.

draw_bboxes = partial(draw_bounding_boxes, fill=False, width=2, font=font_file, font_size=25)

Selecting a Sample Image

We can use the unique ID for an image in the image dictionary to get the image file path and the associated annotations from the annotation DataFrame.

Load the sample image

# Get the file ID of the first image file
file_id = list(img_dict.keys())[0]

# Open the associated image file as a RGB image
sample_img =[file_id]).convert('RGB')

# Print the dimensions of the image
print(f"Image Dims: {sample_img.size}")

# Show the image
Image Dims: (768, 1152)

Inspect the corresponding annotation data

# Get the row from the 'annotation_df' DataFrame corresponding to the 'file_id'
Image Name 258421.jpg
Width 768
Height 1152
Boxes [{‘Label’: ‘person’, ‘xtl’: 386.08, ‘ytl’: 443.94, ‘xbr’: 460.82, ‘ybr’: 777.16}, {‘Label’: ‘person’, ‘xtl’: 340.25, ‘ytl’: 466.94, ‘xbr’: 418.99, ‘ybr’: 777.34}]

Annotate sample image

The draw_bounding_boxes function expects bounding box annotations in [top-left X, top-left Y, bottom-right X, bottom-right Y] format, so we don’t need to convert the annotation values.

# Extract the labels and bounding box annotations for the sample image
labels = [box['Label'] for box in annotation_df.loc[file_id]['Boxes']]
bboxes = np.array([[box['xtl'], box['ytl'], box['xbr'], box['ybr']] for box in annotation_df.loc[file_id]['Boxes']]).reshape(len(labels),4)

# Annotate the sample image with labels and bounding boxes
annotated_tensor = draw_bboxes(
    boxes=BoundingBoxes(torch.Tensor(bboxes), format='xyxy', canvas_size=sample_img.size[::-1]),
    colors=[int_colors[i] for i in [class_names.index(label) for label in labels]]


We have loaded the dataset, inspected its class distribution, and visualized the annotations for a sample image. In the final section, we will cover how to augment images using torchvision’s Transforms V2 API and create a custom Dataset class for training.

Preparing the Data

In this section, we will first walk through a single example of how to apply augmentations to a single annotated image using torchvision’s Transforms V2 API before putting everything together in a custom Dataset class.

Data Augmentation

Here, we will define some data augmentations to apply to images during training. I created a few custom image transforms to help streamline the code.

The first extends the RandomIoUCrop transform included with torchvision to give the user more control over how much it crops into bounding box areas. The second resizes images based on their largest dimension rather than their smallest. The third applies square padding and allows the padding to be applied equally on both sides or randomly split between the two sides.

All three are available through the cjm-torchvision-tfms package.

Set training image size

Next, we will specify the image size to use during training.

# Set training image size
train_sz = 384

Initialize custom transforms

Now, we can initialize the transform objects.

# Create a RandomIoUCrop object
iou_crop = CustomRandomIoUCrop(min_scale=0.3, 
                               sampler_options=[0.0, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0],

# Create a `ResizeMax` object
resize_max = ResizeMax(max_sz=train_sz)

# Create a `PadSquare` object
pad_square = PadSquare(shift=True)

Test the transforms

Torchvision’s V2 image transforms take an image and a targets dictionary. The targets dictionary contains the annotations and labels for the image.

We will pass input through the CustomRandomIoUCrop transform first and then through ResizeMax and PadSquare. We can pass the result through a final resize operation to ensure both sides match the train_sz value.

# Prepare bounding box targets
targets = {
    'boxes': BoundingBoxes(torch.Tensor(bboxes), 
    'labels': torch.Tensor([class_names.index(label) for label in labels])

# Crop the image
cropped_img, targets = iou_crop(sample_img, targets)

# Resize the image
resized_img, targets = resize_max(cropped_img, targets)

# Pad the image
padded_img, targets = pad_square(resized_img, targets)

# Ensure the padded image is the target size
resize = transforms.Resize([train_sz] * 2, antialias=True)
resized_padded_img, targets = resize(padded_img, targets)
sanitized_img, targets = transforms.SanitizeBoundingBoxes()(resized_padded_img, targets)

# Annotate the augmented image with updated labels and bounding boxes
annotated_tensor = draw_bboxes(
    labels=[class_names[int(label.item())] for label in targets['labels']], 
    colors=[int_colors[i] for i in [class_names.index(label) for label in labels]]

# Display the annotated image

    "Source Image:": sample_img.size,
    "Cropped Image:": cropped_img.size,
    "Resized Image:": resized_img.size,
    "Padded Image:": padded_img.size,
    "Resized Padded Image:": resized_padded_img.size,

Source Image: (768, 1152)
Cropped Image: (768, 1152)
Resized Image: (256, 384)
Padded Image: (384, 384)
Resized Padded Image: (384, 384)

Now that we know how to apply data augmentations, we can put all the steps we’ve covered into a custom Dataset class.

Training Dataset Class

The following custom Dataset class is responsible for loading a single image, preparing the associated annotations, applying any image transforms, and returning the final image tensor and its target dictionary during training.

class CVATBBoxDataset(Dataset):
    A custom dataset class for handling bounding box annotations from CVAT.

    This dataset class is designed to work with bounding box annotations exported
    from the CVAT annotation tool. It allows for loading images and their corresponding
    bounding box annotations for use in training machine learning models.

        _img_keys (list): A list of image keys.
        _annotation_df (pandas.DataFrame): A DataFrame containing annotations.
        _img_dict (dict): A dictionary mapping image keys to image file paths.
        _class_to_idx (dict): A dictionary mapping class names to class indices.
        _transforms (callable, optional): A function/transform that takes in an image
            and a target, and returns a transformed version.

        img_keys (list): List of image keys.
        annotation_df (pandas.DataFrame): DataFrame containing annotations.
        img_dict (dict): Dictionary mapping image keys to image paths.
        class_to_idx (dict): Dictionary mapping class names to class indices.
        transforms (callable, optional): Optional transform to be applied on a sample.

    def __init__(self, img_keys, annotation_df, img_dict, class_to_idx, transforms=None):
        super(Dataset, self).__init__()

        self._img_keys = img_keys
        self._annotation_df = annotation_df
        self._img_dict = img_dict
        self._class_to_idx = class_to_idx
        self._transforms = transforms

    def __len__(self):
        Returns the total number of items in the dataset.

            int: Total number of items.
        return len(self._img_keys)

    def __getitem__(self, index):
        Retrieves an image and its associated target (bounding boxes and labels) at the specified index.

            index (int): The index of the item.

            tuple: A tuple containing an image and its corresponding target.
        img_key = self._img_keys[index]
        annotation = self._annotation_df.loc[img_key]
        image, target = self._load_image_and_target(annotation)
        if self._transforms:
            image, target = self._transforms(image, target)

        return image, target

    def _load_image_and_target(self, annotation):
        Loads an image and its corresponding target data (bounding boxes and labels) based on the provided annotation.

            annotation (pandas.Series): The annotation data for a specific image.

            tuple: A tuple containing the image and a dictionary with 'boxes' and 'labels'.
        # Retrieve the file path from the image dictionary using the annotation's name as the key.
        filepath = self._img_dict[]

        # Open the image file and convert it to RGB.
        image ='RGB')

        # Extract bounding box coordinates from the annotation and convert them to a numpy array.
        bbox_list = np.array([[box['xtl'], box['ytl'], box['xbr'], box['ybr']] for box in annotation['Boxes']]).reshape(len(annotation['Boxes']), 4)

        # Convert the numpy array of bounding boxes to a PyTorch tensor.
        bbox_tensor = torch.Tensor(bbox_list)

        # Create bounding box objects with the tensor, specifying the format and canvas size.
        boxes = BoundingBoxes(bbox_tensor, format='xyxy', canvas_size=image.size[::-1])

        # Extract labels from the annotation and map them to their corresponding indices.
        annotation_labels = [box['Label'] for box in annotation['Boxes']]
        labels = torch.Tensor([self._class_to_idx[label] for label in annotation_labels])

        # Return the image and a dictionary containing the bounding boxes and labels.
        return image, {'boxes': boxes, 'labels': labels}

Image Transforms

Here, we will specify and organize all the image transforms to apply during training.

# Compose transforms for data augmentation
data_aug_tfms = transforms.Compose(
                brightness = (0.875, 1.125),
                contrast = (0.5, 1.5),
                saturation = (0.5, 1.5),
                hue = (-0.05, 0.05),
        transforms.RandomPosterize(bits=3, p=0.5),

# Compose transforms to resize and pad input images
resize_pad_tfm = transforms.Compose([
    transforms.Resize([train_sz] * 2, antialias=True)

# Compose transforms to sanitize bounding boxes and normalize input data
final_tfms = transforms.Compose([
    transforms.ToDtype(torch.float32, scale=True),

# Define the transformations for training and validation datasets
train_tfms = transforms.Compose([

Always use the SanitizeBoundingBoxes transform to clean up annotations after using data augmentations that alter bounding boxes (e.g., cropping, warping, etc.).

Initialize Dataset

Now, we can create the dataset object using the image dictionary, the annotation DataFrame, and the image transforms.

# Create a mapping from class names to class indices
class_to_idx = {c: i for i, c in enumerate(class_names)}

# Instantiate the dataset using the defined transformations
train_dataset = CVATBBoxDataset(list(img_dict.keys()), annotation_df, img_dict, class_to_idx, train_tfms)

# Print the number of samples in the training dataset
    'Training dataset size:': len(train_dataset),
Training dataset size: 28

Inspect Samples

To close out, we should verify the dataset object works as intended by inspecting the first sample.

Inspect training set sample

dataset_sample = train_dataset[0]

annotated_tensor = draw_bboxes(
    labels=[class_names[int(i.item())] for i in dataset_sample[1]['labels']], 
    colors=[int_colors[int(i.item())] for i in dataset_sample[1]['labels']]



In this tutorial, we covered how to load custom bounding box annotations made with the CVAT annotation tool and work with them using torchvision’s Transforms V2 API. The skills and knowledge you acquired here provide a solid foundation for future object detection projects.

As a next step, perhaps try annotating a custom object detection dataset with CVAT and loading it with this tutorial’s code. Once you’re comfortable with that, try adapting the code in the following tutorial to train an object detection model on your custom dataset.