Hasty’s online annotation tool is the first to use a human-in-the-loop, AI-assisted approach combined with Active Learning. The result? You can create segmentations 12x quicker than ever before.Start for free Request a demo
With our Semantic Segmentation Assistant, you will get quicker and more effective annotation. Every annotation you make, we send it to our model trainer. We then create a custom model for your project that will speed up your labeling with a factor of 10-100x. This custom-tailored approach to AI annotation means that you can get our AI tools' assistance no matter the use case.
With Hasty, you can directly interact with an AI model - allowing you to debug and de-bias your dataset. Do you see in-tool that the assistant is struggling with labeling lions? Annotate more lion annotations to improve your labeling automation percentage while creating a more robust data asset.
With our models maturing and improving, the need for manual labeling decreases. With enough data, you can batch process the rest, or a specific part, of your dataset in one click.
How do we pick the best image to label next? Usually, we grab the next one in the queue. But with Hasty's new Active Learning feature, we use state-of-the-art machine learning to build a labeling queue for you and your labelers, where we create an optimal labeling queue. The result? You never waste time labeling what your model already understands while getting better-performing models faster.
Using our Model Playground, you can replace the labeling automation models we train for you with your own. With this approach, you can create models achieving SOTA performance while, at the same time, being able to deploy them for better labeling automation in one click.
The difference in Hasty is that everything is custom. For every project, we train custom models to assist you in labeling faster. Doing so means we can provide you with higher and higher degrees of automation as you label more and more data.
For example, we recently had a customer who spent 22 minutes on average labeling an image using semantic segmentation. With Hasty's automation, we cut that down to 15 minutes in week 1, 12 minutes in week 2, and 1 minute in week 4 - making them 95% faster labeling.
In Hasty, labeling is completely free*. We also don't have any usage or user limits for our free tier. In comparison, most of our competitors will charge you thousands, or in some cases, tens of thousands of dollars per year for worse tooling.
* Our batch processing feature is not free, but all other semantic segmentation automation features are.
We want to say it's just us being nice. However, the truth is that if we can give you a great user experience and provide high degrees of labeling automation, you will be able to build a substantial data asset.
With such a data asset, you can get a significant boost to QA using our premium AI Consensus Scoring feature, and you would be more likely to experiment with our Model Playground. These are paid-for features, and that's how we plan to make money.
In short, we take your data and automatically train a model.
Then, we make this model available through our Semantic Segmentation Assistant when you label. We retrain the model every time you add 20% more data (going from 10 to 12 to 15 images, for example). Doing so means you will get better labeling automation the more you use the platform.
We think our approach has a couple of advantages compared with other techniques:
- As our automation trains on your data, we can guarantee that you get labeling automation working on your use case, whatever it is.
- Because we allow you to accept and reject suggestions, you avoid common problems with pre-labeling approaches where you have to spend as much time editing and fixing images as it would take labeling them from scratch
- You also get constant feedback on what the model understands and where it struggles. With this, you can adapt your data strategy, giving the model more examples of data it struggles with. Or you can train a better model using Model Playground.
We currently use an FPN architecture using a ConvNeXt backbone. This combination gives a sound basis for all use cases and allows us to deliver exemplary levels of automation as-is.
However, if you want to use more state-of-the-art semantic segmentation neural networks for better automation, you can use Model Playground. For example, we just added Mask2Former, one of the best-performing architectures for semantic segmentation as of 2022.
For now, 2D images (of any well-used format). We support 8K resolutions and can handle hundreds of labels per image if you have the hardware to support it.
We support it, although it requires a bit of a particular setup. Contact us if you are interested in trying Hasty to do semantic segmentation on spatial data.
Yes, you can. Our import function currently supports COCO, Pascal VOC, and semantic PNGs formats but we can always help if you need to import another format into Hasty (just send a request here).
Although we see ourselves primarily as a web-based, online computer vision platform, we have some customers for whom on-premise or hybrid deployments are required and can support deployment to most common environments.
If you are interested in these deployments, contact us here to learn more.
No. Although we train models on your data, we only make those models available for you. We never use your data to train models for anyone but you. We also never look at your data without written permission to do so. All of this is in our T&Cs, and as such, legally binding.
Yes. We can mount storage buckets from Azure, GCP, and AWS and fetch images from your storage on demand instead of storing them on our servers.
Since October of 2022, we do have semi-supervised functionality available in Hasty.
We primarily recommend using it in cases where you have a small amount of labeled data, and a large amount of unlabelled data. If you fulfill those requirements and want to know more, feel free to book a demo here.
Yes. We are happy to offer guidance for free to make sure you are successful using Hasty. Book a demo and we'll do our best assisting you.
Our servers are located in Belgium.
Yes. Having been acquired by CloudFactory, which is one of the world's largest workforce providers for machine learning data creation, we can provide you with a combined offering for software and workforce. If you are interested, feel free to make an inquiry here.
Hasty is based in Berlin, Germany.
Dr. Alexander Roth