Creating Augmented Images

Creating augmented images is a critical step to enhance model performance by improving its ability to generalize to unseen data.

Image augmentation involves applying transformations to the existing images in your dataset. This process can help your model generalize better and perform more effectively on new, unseen images.

Supported Augmentations in BinaExperts:

· Flip

· 90-degree rotation

· Random rotation

· Random crop

· Random shear

· Blur

· Exposure adjustment

· Random noise

· Cutout (available in paid plans only)

· Mosaic (available in paid plans only)

Recommendations for Starting a Project:

We recommend beginning your project without any augmentations. This approach helps you evaluate the quality of your raw dataset. If your dataset's performance is suboptimal with augmentations, you will not have a clear baseline for comparison.

If your model does not perform well without augmentations, you should first investigate factors like class balance, data representation, and dataset size. Once your model performs satisfactorily on the raw dataset, you can add augmentations to further enhance its performance.

Creating Augmented Images Before Training

Conducting augmentations offline (prior to training) offers several benefits:

1. Improved Model Reproducibility: BinaExperts keeps a record of how each image was augmented, allowing you to trace back augmentation settings. For instance, if your model performs better on brighter images, you can identify the need for more low-light training data.

2. Reduced Training Time: Augmentations are CPU-intensive operations. When augmentations are performed on-the-fly during training, your GPU may often wait for your CPU to generate augmented data for each epoch, slowing down the process.

3. Lower Training Costs: As augmentations are CPU-bound, performing them offline can reduce the waiting time for your GPU, making your training process more efficient and cost-effective.

Adding Augmentations

To add augmentations, navigate to the "Versions" tab in your project’s dashboard on BinaExperts. Click on "Augmentations" to set up your desired augmentation parameters.

You can specify how many times each image should be augmented. For instance, selecting "3" means that each image will be augmented twice based on your settings, resulting in three versions of each source image.

How Augmentations Are Applied

Augmentations are chained together with random settings applied to each image. Duplicates generated during this process are filtered out.

For example, if you choose "flip horizontally" and "add random noise," an image may be randomly flipped and then have noise applied.

Augmentation Options and Customization

Below are the available augmentations in BinaExperts and their customizable parameters:

available augmentations in BinaExperts and their customizable parameters

1. Flip: Randomly flip images vertically or horizontally with correctly mirrored annotations.

· Horizontal: Flips the image left-to-right.

· Vertical: Flips the image top-to-bottom.

1. 90-Degree Rotations: Rotate images by 90 or 180 degrees.

· Clockwise: Rotates 90 degrees clockwise.

· Counter Clockwise: Rotates 90 degrees counter-clockwise.

· Upside Down: Rotates 180 degrees.

1. Random Rotation: Rotate images randomly up to a specified degree. Note that annotations may be cropped if they are partially out of frame.

2. Random Shear: Distorts images horizontally or vertically.

· Horizontal: Maximum shear along the x-axis.

· Vertical: Maximum shear along the y-axis.

1. Blur: Applies Gaussian blur to the image.

· Pixels: Determines the intensity of blur (maximum of 25 pixels).

1. Random Noise: Adds salt-and-pepper noise to the image.

· Percent: The percentage of pixels affected (up to 25%).

7. Color Adjustments:

· Hue, Value, Saturation (HVS): Modifies the color appearance of images.

· Contrast: Alters the difference between light and dark areas.

1. Weather Effects:

· Snow: Simulates snow-like conditions.

· Fog: Adds foggy or misty effects.

· Rain: Simulates raindrop effects.

· Sun: Introduces sunlight effects, altering brightness and warmth.

These augmentation options allow you to generate diverse variations in your dataset, enhancing model robustness and improving generalization to real-world scenarios.