Synthetic data generation in Supervisely turns a few labeled objects into new training data. Apps paste augmented copies of your objects onto backgrounds, so every generated image or video is labeled for detection, segmentation or tracking.
Labeling is usually the slowest part of training a model, and every new scenario adds more of it. Synthetic data turns most of that work into settings you can change and run again.
The app places each object itself, so it knows every label exactly. People label a few examples instead of every image.
People outline every object on every image. It costs time and money, and people make mistakes.
Add backgrounds, change object counts or augmentations, and generate the dataset again.
New conditions, new objects or a new camera send the team back to labeling.
If something is off, correct the settings and regenerate the whole dataset.
An error found late means finding and correcting labels image by image.
The real world cannot be simulated perfectly. It works better to randomize backgrounds, positions and colors so that the model learns to adapt, an approach known as domain randomization.
What we learned training on synthetic dataSynthetic data does not suit every task. Tell us about yours, and our team will help you find out.
Talk to an expertAn example from our team: an instance segmentation dataset of seeds, generated with the Flying objects app. The same pipeline works for other tasks, such as microbe or pathology detection.
Outline some objects with polygons or masks, or take labeled objects from another source. Typical backgrounds are easy to find or generate. Under 70 photos covered every case here.
objects:
count: [200, 300] # objects per image
visibility: 0.8 # share left uncovered
augs:
color:
RandomBrightnessContrast: true
CLAHE: true
Blur: trueRun Flying objects from the project’s context menu, pick the classes and background images, set augmentations in YAML and preview a random result.
The app saves the new images with their labels to a new or existing project. Check the results and start training.
Flying objects makes synthetic training data for object detection, semantic segmentation and instance segmentation, and checks every object it pastes.
Every object stays at least 80% visible by default, a threshold you can change. Objects that cannot be placed this way are skipped, and hidden parts are cut from the labels.
Paste objects labeled with polygons or masks, the shapes that let the app check overlaps. Get one mask per class, bounding boxes or a mask for every object.
Every pasted object gets its own color augmentations from Albumentations, flips, rotation and resizing from ImgAug, and soft edges and opacity that blend it into the background.
Take backgrounds from any project or dataset. Their polygon and mask labels of the selected classes can be kept and merged with the new ones, which makes it copy-paste augmentation for real labeled data.
Check a random preview, then generate up to 1,000 images per run into a new or existing project, and keep each experiment in its own dataset.
objects: # objects per image, for every class count: [200, 300] # min visible share of each object visibility: 0.8 # soft edges and transparency edge_smoothing_ksize: [5, 9] opacity: [0.8, 1.0] augs: color: # Albumentations RandomBrightnessContrast: true CLAHE: true Blur: true spatial: # ImgAug Fliplr: 0.5 Flipud: 0.5 Rotate: (-90, 90) Resize: (0.8, 1.5)
Two more apps make data for product recognition and object tracking, and our research on cracks shows how far synthetic data can go.
Synthetic retail products turns one tagged example per product into training and validation images, with random background colors, rotation, elastic distortion and other products partly covering it. Each product gets its own dataset of up to 10,000 training images.
Synthetic retail productsSynthetic videos for tracking moves labeled objects over a fixed background along straight or random paths, and saves every video with a bounding box track for each object. You set the speed, the allowed overlap, 10 to 60 FPS and the duration.
Synthetic videos for trackingFor thin, branching objects, our team generated cracks procedurally and drew them onto about 200 photos of asphalt, concrete, walls and other materials. The result is the Cracks Synthetic Dataset in the Ecosystem.
Cracks Synthetic DatasetA model trained only on our synthetic cracks came close to one trained on 628 hand-labeled photos. With 628 real photos added without labels through HRDA domain adaptation, a larger model trained this way went past that reference.
Synthetic images are saved to a regular project, so the labeling, data and model tools of the platform work on them as on real data. See the platform overview.
Polygons, brushes and the Smart Tool, which turns a box and a few clicks into a precise mask.
Filter images, convert shapes and apply classic augmentations to whole datasets in ML Pipelines.
Train detection and segmentation models on the generated project, then put them to work.
Evaluate models on real labeled images, with reports that compare checkpoints.
It is making training images with a program instead of photographing and labeling them by hand. Because the program places every object itself, each generated image comes with exact labels. In Supervisely, apps do this by pasting augmented copies of labeled objects onto background images.
Object detection, semantic segmentation and instance segmentation with the Flying objects app, classification and segmentation of retail products with Synthetic retail products, and object tracking in video with Synthetic videos for tracking. Each app runs from the context menu of an images project.
Copy-paste augmentation cuts labeled objects out of some images and pastes them into others, which gives new training examples with known labels. Flying objects pastes objects labeled with polygons or masks onto empty backgrounds or onto labeled images, and can keep the labels those images already have.
Augmentation changes images you already have, for example by flipping, cropping or blurring them. Synthetic data generation composes new images from labeled objects and backgrounds. The two work together: Flying objects augments every object it pastes, and ML Pipelines augment whole datasets.
Far fewer than for a dataset labeled by hand. In our seeds example, under 70 photos covered every case. Label the objects with polygons or masks: these are the shapes that let the app check overlaps and keep every label exact.
It can, if the synthetic data is varied enough, and it should always be checked on real labeled images. In our crack segmentation research, a model trained only on synthetic images came close to one trained on 628 hand-labeled photos. With HRDA domain adaptation and 628 real photos added without labels, a larger model went past it. Read the research.