Image Transforms

Image transforms are essential for nearly every computer vision task. They are used for preprocessing (e.g., resizing images to a uniform size, normalizing pixel values) and, critically, for data augmentation (e.g., applying random rotations, crops, and color shifts to the training data). Data augmentation is a key technique for preventing overfitting and improving model robustness.

xTorch provides a massive library of image transforms, from basic operations to advanced, state-of-the-art augmentation techniques.

All image transforms are located under the xt::transforms::image namespace and their headers can be found in the <xtorch/transforms/image/> directory.

General Usage

The most common way to use image transforms is to chain them together into a pipeline using xt::transforms::Compose. This pipeline is then passed to your Dataset during construction, which will apply the transformations to each image as it is loaded.

#include <xtorch/xtorch.h>
#include <iostream>
 
int main() {
    // This is a standard data augmentation pipeline for training on ImageNet
    auto training_transforms = std::make_unique<xt::transforms::Compose>(
        // Resize the smaller edge to 256, maintaining aspect ratio
        std::make_shared<xt::transforms::image::Resize>(256),
        // Randomly crop a 224x224 patch
        std::make_shared<xt::transforms::image::RandomCrop>(std::vector<int64_t>{224, 224}),
        // Randomly flip the image horizontally with a 50% probability
        std::make_shared<xt::transforms::image::RandomHorizontalFlip>(/*p=*/0.5),
        // Apply some color jitter
        std::make_shared<xt::transforms::image::ColorJitter>(),
        // Normalize the image using ImageNet mean and stddev
        std::make_shared<xt::transforms::general::Normalize>(
            std::vector<float>{0.485, 0.456, 0.406},
            std::vector<float>{0.229, 0.224, 0.225}
        )
    );
 
    // This pipeline would be passed to a dataset
    // auto dataset = xt::datasets::ImageFolderDataset("./data", std::move(training_transforms));
}

!!! info "Constructor Options" Nearly all transforms are configurable through their constructors. This includes parameters like sizes, probabilities (p), rotation degrees, and more. Always refer to the specific header file in <xtorch/transforms/image/> for a full list of available settings.

Available Transforms by Category

Geometric Transforms

These transforms alter the spatial properties of the image.

Color & Photometric Transforms

These transforms alter the pixel values, colors, brightness, and contrast of the image.

Augmentation & Erasing Transforms

These are advanced augmentation techniques that often involve erasing or mixing parts of images.

Blur & Noise Transforms

| GlassBlur | Applies a glass-like blur effect. | | ZoomBlur | Applies a blur that simulates zooming. | | GaussianNoise| Adds Gaussian noise to the image. | | NoiseInjection| Injects random noise into the image. |

Stylistic & Filter-Based Transforms