Learn how Supervisely can dramatically improve model training results or even make impossible — possible!
A traditional approach to model training has many steps, but it always includes the most challenging part: data annotation. There are many potential issues:
Fortunately, there is a better way. Instead of labeling dataset, we can programmatically compose it from different parts. For example, we are building an OCR system for shop receipts. Rather than labeling photos of read-world receipts, we can generate infinite amount of synthetic receipts that look almost identical. Since we generate all the labels, we don't need the labeling and can proceed straight to the model building.
Almost identical? This may seem like a problem at first, but many researches show that it's impossible to perfectly simulate real world — instead, it's better to teach model to adapt and learn from randomization.
Synthetic data generation approach doesn't applicable for every task. But, usually, when it works, it works very well. What about you case? Get in touch with us and we will figure it out!
Here is an example of generating a synthetic dataset for instance segmentation of seeds. Exactly the same pipeline works for many other tasks, such as microbes or pathology detection.
Label subset of you data or obtain it from external sources
Usually, it's easy to find or generate typical backgrounds
Just under 70 images is enough to represent every case we want to cover
Let's apply synthetic generation application from the Ecosystem with an appropriate configuration
Generate synthetic data: flying foregrounds on top of backgrounds
Done! Verify generated data and start model training
Because Supervisely is built like OS for computer vision, we made possible integration of the best machine learning models and tools on a single platform.
You will find a well-known projects from data science community, as well as our own Apps, providing a complete solution for entire AI development pipeline.
Configure every aspect of training from target classes to online augmentations, monitor metrics and terminal logs in real-time.
Dashboard to configure and monitor training
Understand how your model works on ground truth and new data and find how to correct negative output and increase performance.
Detailed statistics for all classes in images project
Generate synthetic datasets that drastically improve model results, especially when there is not enough ground truth.
Generate synthetic data: flying foregrounds on top of backgrounds
Perform all the necessary actions on your data, from importing and converting to skeletonization of masks and rasterization.
Configure, preview and split images and annotations with sliding window
A fully customizable AI infrastructure, deployed on cloud or your servers with everything you love about Supervisely, plus advanced security, control, and support.
Start 30 days free trial➔Supervisely provides first-rate experience since 2017, longer than most of computer vision platforms over there.
Join community of thousands computer vision enthusiasts and companies of every size that use Supervisely every day.
Our online version has over a 220 million of images and over a billion of labels created by our great community.
Speak with people who are on the same page with you. An actual data scientist will:
Get accurate training data on scale with expert annotators, ML-assisted tools, dedicated project manager and the leading labeling platform.
Order workforce