Data augmentation is a crucial technique in computer vision that artificially expands your training dataset by applying various transformations to existing images. When training deep learning models like Ultralytics YOLO, data augmentation helps improve model robustness, reduces overfitting, and enhances generalization to real-world scenarios.
Why Data Augmentation Matters
Data augmentation serves multiple critical purposes in training computer vision models:
Expanded Dataset: By creating variations of existing images, you can effectively increase your training dataset size without collecting new data.
Improved Generalization: Models learn to recognize objects under various conditions, making them more robust in real-world applications.
Reduced Overfitting: By introducing variability in the training data, models are less likely to memorize specific image characteristics.
Enhanced Performance: Models trained with proper augmentation typically achieve better accuracy on validation and test sets.
Ultralytics YOLO's implementation provides a comprehensive suite of augmentation techniques, each serving specific purposes and contributing to model performance in different ways. This guide will explore each augmentation parameter in detail, helping you understand when and how to use them effectively in your projects.
Example Configurations
You can customize each parameter using the Python API, the command line interface (CLI), or a configuration file. Below are examples of how to set up data augmentation in each method.
Configuration Examples
fromultralyticsimportYOLO# Load a modelmodel=YOLO("yolo11n.pt")# Training with custom augmentation parametersmodel.train(data="coco.yaml",epochs=100,hsv_h=0.03,hsv_s=0.6,hsv_v=0.5)# Training without any augmentations (disabled values omitted for clarity)model.train(data="coco.yaml",epochs=100,hsv_h=0.0,hsv_s=0.0,hsv_v=0.0,translate=0.0,scale=0.0,fliplr=0.0,mosaic=0.0,erasing=0.0,auto_augment=None,)
# Training with custom augmentation parametersyolodetecttraindata=coco8.yamlmodel=yolo11n.ptepochs=100hsv_h=0.03hsv_s=0.6hsv_v=0.5
Using a configuration file
You can define all training parameters, including augmentations, in a YAML configuration file (e.g., train_custom.yaml). The mode parameter is only required when using the CLI. This new YAML file will then override the default one located in the ultralytics package.
# train_custom.yaml# 'mode' is required only for CLI usagemode:traindata:coco8.yamlmodel:yolo11n.ptepochs:100hsv_h:0.03hsv_s:0.6hsv_v:0.5
Then launch the training with the Python API:
Train Example
fromultralyticsimportYOLO# Load a COCO-pretrained YOLO11n modelmodel=YOLO("yolo11n.pt")# Train the model with custom configurationmodel.train(cfg="train_custom.yaml")
# Train the model with custom configurationyolodetecttrainmodel="yolo11n.pt"cfg=train_custom.yaml
Color Space Augmentations
Hue Adjustment (hsv_h)
Range: 0.0 - 1.0
Default: 0.015
Usage: Shifts image colors while preserving their relationships. The hsv_h hyperparameter defines the shift magnitude, with the final adjustment randomly chosen between -hsv_h and hsv_h. For example, with hsv_h=0.3, the shift is randomly selected within-0.3 to 0.3. For values above 0.5, the hue shift wraps around the color wheel, that's why the augmentations look the same between 0.5 and -0.5.
Purpose: Particularly useful for outdoor scenarios where lighting conditions can dramatically affect object appearance. For example, a banana might look more yellow under bright sunlight but more greenish indoors.
Usage: Modifies the intensity of colors in the image. The hsv_h hyperparameter defines the shift magnitude, with the final adjustment randomly chosen between -hsv_s and hsv_s. For example, with hsv_s=0.7, the intensity is randomly selected within-0.7 to 0.7.
Purpose: Helps models handle varying weather conditions and camera settings. For example, a red traffic sign might appear highly vivid on a sunny day but look dull and faded in foggy conditions.
Usage: Changes the brightness of the image. The hsv_v hyperparameter defines the shift magnitude, with the final adjustment randomly chosen between -hsv_v and hsv_v. For example, with hsv_v=0.4, the intensity is randomly selected within-0.4 to 0.4.
Purpose: Essential for training models that need to perform in different lighting conditions. For example, a red apple might look bright in sunlight but much darker in the shade.
Usage: Rotates images randomly within the specified range. The degrees hyperparameter defines the rotation angle, with the final adjustment randomly chosen between -degrees and degrees. For example, with degrees=10.0, the rotation is randomly selected within-10.0 to 10.0.
Purpose: Crucial for applications where objects can appear at different orientations. For example, in aerial drone imagery, vehicles can be oriented in any direction, requiring models to recognize objects regardless of their rotation.
Usage: Shifts images horizontally and vertically by a random fraction of the image size. The translate hyperparameter defines the shift magnitude, with the final adjustment randomly chosen twice (once for each axis) within the range -translate and translate. For example, with translate=0.5, the translation is randomly selected within-0.5 to 0.5 on the x-asis, and another independent random value is selected within the same range on the y-axis.
Purpose: Helps models learn to detect partially visible objects and improves robustness to object position. For example, in vehicle damage assessment applications, car parts may appear fully or partially in frame depending on the photographer's position and distance, the translation augmentation will teach the model to recognize these features regardless of their completeness or position.
Note: For simplicity, the translations applied below are the same each time for both x and y axes. Values -1.0 and 1.0are not shown as they would translate the image completely out of the frame.
-0.5
-0.25
0.0
0.25
0.5
Scale (scale)
Range: ≥0.0
Default: 0.5
Usage: Resizes images by a random factor within the specified range. The scale hyperparameter defines the scaling factor, with the final adjustment randomly chosen between 1-scale and 1+scale. For example, with scale=0.5, the scaling is randomly selected within0.5 to 1.5.
Purpose: Enables models to handle objects at different distances and sizes. For example, in autonomous driving applications, vehicles can appear at various distances from the camera, requiring the model to recognize them regardless of their size.
The value -1.0 is not shown as it would make the image disappear, while 1.0 simply results in a 2x zoom.
The values displayed in the table below are the ones applied through the hyperparameter scale, not the final scale factor.
If scale is greater than 1.0, the image can be either very small or flipped, as the scaling factor is randomly chosen between 1-scale and 1+scale. For example, with scale=3.0, the scaling is randomly selected within-2.0 to 4.0. If a negative value is chosen, the image is flipped.
-0.5
-0.25
0.0
0.25
0.5
Shear (shear)
Range: -180 to +180
Default: 0.0
Usage: Introduces a geometric transformation that skews the image along both x-axis and y-axis, effectively shifting parts of the image in one direction while maintaining parallel lines. The shear hyperparameter defines the shear angle, with the final adjustment randomly chosen between -shear and shear. For example, with shear=10.0, the shear is randomly selected within-10 to 10 on the x-asis, and another independent random value is selected within the same range on the y-axis.
Purpose: Helps models generalize to variations in viewing angles caused by slight tilts or oblique viewpoints. For instance, in traffic monitoring, objects like cars and road signs may appear slanted due to non-perpendicular camera placements. Applying shear augmentation ensures the model learns to recognize objects despite such skewed distortions.
shear values can rapidly distort the image, so it's recommended to start with small values and gradually increase them.
Unlike perspective transformations, shear does not introduce depth or vanishing points but instead distorts the shape of objects by changing their angles while keeping opposite sides parallel.
-10
-5
0.0
5
10
Perspective (perspective)
Range: 0.0 - 0.001
Default: 0.0
Usage: Applies a full perspective transformation along both x-axis and y-axis, simulating how objects appear when viewed from different depths or angles. The perspective hyperparameter defines the perspective magnitude, with the final adjustment randomly chosen between -perspective and perspective. For example, with perspective=0.001, the perspective is randomly selected within-0.001 to 0.001 on the x-asis, and another independent random value is selected within the same range on the y-axis.
Purpose: Perspective augmentation is crucial for handling extreme viewpoint changes, especially in scenarios where objects appear foreshortened or distorted due to perspective shifts. For example, in drone-based object detection, buildings, roads, and vehicles can appear stretched or compressed depending on the drone's tilt and altitude. By applying perspective transformations, models learn to recognize objects despite these perspective-induced distortions, improving their robustness in real-world deployments.
Usage: Performs a vertical flip by inverting the image along the y-axis. This transformation mirrors the entire image upside-down but preserves all spatial relationships between objects. The flipud hyperparameter defines the probability of applying the transformation, with a value of flipud=1.0 ensuring that all images are flipped and a value of flipud=0.0 disabling the transformation entirely. For example, with flipud=0.5, each image has a 50% chance of being flipped upside-down.
Purpose: Useful for scenarios where objects can appear upside down. For example, in robotic vision systems, objects on conveyor belts or robotic arms may be picked up and placed in various orientations. Vertical flipping helps the model recognize objects regardless of their top-down positioning.
Usage: Performs a horizontal flip by mirroring the image along the x-axis. This transformation swaps the left and right sides while maintaining spatial consistency, which helps the model generalize to objects appearing in mirrored orientations. The fliplr hyperparameter defines the probability of applying the transformation, with a value of fliplr=1.0 ensuring that all images are flipped and a value of fliplr=0.0 disabling the transformation entirely. For example, with fliplr=0.5, each image has a 50% chance of being flipped left to right.
Purpose: Horizontal flipping is widely used in object detection, pose estimation, and facial recognition to improve robustness against left-right variations. For example, in autonomous driving, vehicles and pedestrians can appear on either side of the road, and horizontal flipping helps the model recognize them equally well in both orientations.
Usage: Swaps the color channels of an image from RGB to BGR, altering the order in which colors are represented. The bgr hyperparameter defines the probability of applying the transformation, with bgr=1.0 ensuring all images undergo the channel swap and bgr=0.0 disabling it. For example, with bgr=0.5, each image has a 50% chance of being converted from RGB to BGR.
Purpose: Increases robustness to different color channel orderings. For example, when training models that must work across various camera systems and imaging libraries where RGB and BGR formats may be inconsistently used, or when deploying models to environments where the input color format might differ from the training data.
Usage: Combines four training images into one. The mosaic hyperparameter defines the probability of applying the transformation, with mosaic=1.0 ensuring that all images are combined and mosaic=0.0 disabling the transformation. For example, with mosaic=0.5, each image has a 50% chance of being combined with three other images.
Purpose: Highly effective for improving small object detection and context understanding. For example, in wildlife conservation projects where animals may appear at various distances and scales, mosaic augmentation helps the model learn to recognize the same species across different sizes, partial occlusions, and environmental contexts by artificially creating diverse training samples from limited data.
Even if the mosaic augmentation makes the model more robust, it can also make the training process more challenging.
The mosaic augmentation can be disabled near the end of training by setting close_mosaic to the number of epochs before completion when it should be turned off. For example, if epochs is set to 200 and close_mosaic is set to 20, the mosaic augmentation will be disabled after 180 epochs. If close_mosaic is set to 0, the mosaic augmentation will be enabled for the entire training process.
The center of the generated mosaic is determined using random values, and can either be inside the image or outside of it.
The current implementation of the mosaic augmentation combines 4 images picked randomly from the dataset. If the dataset is small, the same image may be used multiple times in the same mosaic.
mosaic off
mosaic on
Mixup (mixup)
Range: 0.0 - 1.0
Default: 0.0
Usage: Blends two images and their labels with given probability. The mixup hyperparameter defines the probability of applying the transformation, with mixup=1.0 ensuring that all images are mixed and mixup=0.0 disabling the transformation. For example, with mixup=0.5, each image has a 50% chance of being mixed with another image.
Purpose: Improves model robustness and reduces overfitting. For example, in retail product recognition systems, mixup helps the model learn more robust features by blending images of different products, teaching it to identify items even when they're partially visible or obscured by other products on crowded store shelves.
The mixup ratio is a random value picked from a np.random.beta(32.0, 32.0) beta distribution, meaning each image contributes approximately 50%, with slight variations.
First image, mixup off
Second image, mixup off
mixup on
Segmentation-Specific Augmentations
Copy-Paste (copy_paste)
Range: 0.0 - 1.0
Default: 0.0
Usage: Only works for segmentation tasks, this augmentation copies objects within or between images based on a specified probability, controlled by the copy_paste_mode. The copy_paste hyperparameter defines the probability of applying the transformation, with copy_paste=1.0 ensuring that all images are copied and copy_paste=0.0 disabling the transformation. For example, with copy_paste=0.5, each image has a 50% chance of having objects copied from another image.
Purpose: Particularly useful for instance segmentation tasks and rare object classes. For example, in industrial defect detection where certain types of defects appear infrequently, copy-paste augmentation can artificially increase the occurrence of these rare defects by copying them from one image to another, helping the model better learn these underrepresented cases without requiring additional defective samples.
As pictured in the gif below, the copy_paste augmentation can be used to copy objects from one image to another.
Once an object is copied, regardless of the copy_paste_mode, its Intersection over Area (IoA) is computed with all the object of the source image. If all the IoA are below a 0.3, the object is pasted in the target image. If only one the IoA is above 0.3, the object is not pasted in the target image.
The IoA threshold cannot be changed with the current implementation and is set to 0.3 by default.
copy_paste off
copy_paste on with copy_paste_mode=flip
Visualize the copy_paste process
Copy-Paste Mode (copy_paste_mode)
Options: 'flip', 'mixup'
Default: 'flip'
Usage: Determines the method used for copy-paste augmentation. If set to 'flip', the objects come from the same image, while 'mixup' allows objects to be copied from different images.
Purpose: Allows flexibility in how copied objects are integrated into target images.
Usage: Applies automated augmentation policies for classification. The 'randaugment' option uses RandAugment, 'autoaugment' uses AutoAugment, and 'augmix' uses AugMix. Setting to None disables automated augmentation.
Purpose: Optimizes augmentation strategies automatically for classification tasks. The differences are the following:
AutoAugment: This mode applies predefined augmentation policies learned from datasets like ImageNet, CIFAR10, and SVHN. Users can select these existing policies but cannot train new ones within Torchvision. To discover optimal augmentation strategies for specific datasets, external libraries or custom implementations would be necessary. Reference to the AutoAugment paper.
RandAugment: Applies a random selection of transformations with uniform magnitude. This approach reduces the need for an extensive search phase, making it more computationally efficient while still enhancing model robustness. Reference to the RandAugment paper.
AugMix: AugMix is a data augmentation method that enhances model robustness by creating diverse image variations through random combinations of simple transformations. Reference to the AugMix paper.
Essentially, the main difference between the three methods is the way the augmentation policies are defined and applied.
You can refer to this article that compares the three methods in detail.
Random Erasing (erasing)
Range: 0.0 - 0.9
Default: 0.4
Usage: Randomly erases portions of the image during classification training. The erasing hyperparameter defines the probability of applying the transformation, with erasing=0.9 ensuring that almost all images are erased and erasing=0.0 disabling the transformation. For example, with erasing=0.5, each image has a 50% chance of having a portion erased.
Purpose: Helps models learn robust features and prevents over-reliance on specific image regions. For example, in facial recognition systems, random erasing helps models become more robust to partial occlusions like sunglasses, face masks, or other objects that might partially cover facial features. This improves real-world performance by forcing the model to identify individuals using multiple facial characteristics rather than depending solely on distinctive features that might be obscured.
The erasing augmentation comes with a scale, ratio, and value hyperparameters that cannot be changed with the current implementation. Their default values are (0.02, 0.33), (0.3, 3.3), and 0, respectively, as stated in the PyTorch documentation.
The upper limit of the erasing hyperparameter is set to 0.9 to avoid applying the transformation to all images.
erasing off
erasing on (example 1)
erasing on (example 2)
erasing on (example 3)
FAQ
There are too many augmentations to choose from. How do I know which ones to use?
Choosing the right augmentations depends on your specific use case and dataset. Here are a few general guidelines to help you decide:
In most cases, slight variations in color and brightness are beneficial. The default values for hsv_h, hsv_s, and hsv_v are a solid starting point.
If the camera's point of view is consistent and won't change once the model is deployed, you can likely skip geometric transformations such as rotation, translation, scale, shear, or perspective. However, if the camera angle may vary, and you need the model to be more robust, it's better to keep these augmentations.
Use the mosaic augmentation only if having partially occluded objects or multiple objects per image is acceptable and does not change the label value. Alternatively, you can keep mosaic active but increase the close_mosaic value to disable it earlier in the training process.
In short: keep it simple. Start with a small set of augmentations and gradually add more as needed. The goal is to improve the model's generalization and robustness, not to overcomplicate the training process. Also, make sure the augmentations you apply reflect the same data distribution your model will encounter in production.
When starting a training, a see a albumentations: Blur[...] reference. Does that mean Ultralytics YOLO runs additional augmentation like blurring?
If the albumentations package is installed, Ultralytics automatically applies a set of extra image augmentations using it. These augmentations are handled internally and require no additional configuration.
You can find the full list of applied transformations in our technical documentation, as well as in our Albumentations integration guide. Note that only the augmentations with a probability p greater than 0 are active. These are purposefully applied at low frequencies to mimic real-world visual artifacts, such as blur or grayscale effects.
When starting a training, I don't see any reference to albumentations. Why?
Check if the albumentations package is installed. If not, you can install it by running pip install albumentations. Once installed, the package should be automatically detected and used by Ultralytics.
