In most datasets, males, light-skinned people, and adults aged 18-40 are overrepresented. Bias is bound to occur in any real-life data collection method because of geographic constraints. It’s simply too time consuming to legally collect enough data at scale that balances permutations of skin tone, sexes, and age–not to mention camera position and lighting variations. Our datasets help fill in the gaps and re-balance your model.
Our synthesized identities represent people from all around the world, including multi-racial ethnicities. Pick and choose the ones you need to balance your existing model, or get a fully balanced set from us, ready-to-go.
Procedural Facial Attributes
Even in 2020, the largest providers of Vision APIs like Google, Microsoft, and IBM still have significant sex bias in their results*. Easily generate fully-accessorized males and females with our programmatic APIs, or download a pre-made dataset to make your models more fair. * According to 2020 research by Wunderman Thompson
Our datasets feature adults 18 and beyond, in an even distribution in combination with other attributes. With our API, you can fine-tune our images to exactly what your model needs to reduce its bias.
Camera Angle & Pose
Publically available datasets primarily feature subjects from a frontal viewpoint, like pictures you might see taken on the red carpet. But in most situations, either the camera isn’t head-on, or people aren’t still and posing for a shot. Our datasets provide camera angles and head poses from every-which-direction.
Lighting & Environment
Pictures taken by real people in real situations are rarely as well lit as a studio or have professional flash hardware available. With fully customizable environments and lights, you can be sure your model robustly handles any shot.
Synthesizing the Solution
With programmatic creation, we provide vast amounts of diverse training data for better-generalized computer vision models; all while maintaining user privacy.
The average amount spent on single image for full-segmentation is $6.40* – any additional labels cost more above and beyond that. Our synthetic data provides full-segmentation, landmarks, surface normals, and more – for as little as $0.03 per image.
Of course, that’s only the labeling cost. Procuring the images to label is incredibly time-consuming as well. It can take weeks or months to legally collect diverse images of individuals’ faces for most companies. Our datasets are available immediately, and our programmatic API returns generated images and labels in minutes to hours.
*Based on scale.ai pricing, January 2021.
Not only do our datasets provide training data affordably and nearly instantaneously, they do so much more than human-collected & labeled data. So you can build more advanced, more ethical computer vision models.
100% accurate ground truth–every time. Eliminate your QA step on every label.
Get peace of mind: with non-real humans, privacy concerns are history.
Even sampling across skin tones, ethnicities, and ages, for more ethical machine vision.
New Label Types
Use cutting edge models with depth, normals, dense 3D landmarks, & subsegmentation.
Combine identities, hair styles, facial hair, makeup, hats, glasses, face masks, lighting conditions, and camera angles for trillions of possibilities – all at the speed of writing JSON with our API.
Get Going Fast
Check out our snippets to jump-start the training process.
As with all synthetic data, there’s a shift from our domain to the one captured by real cameras. Although there’s no universal domain adaptation approach for every use-case, we stand on the shoulders of giants to get great results.
Adaptive Batch Normalization
Adaptive Batch Normalization is a simple technique, can be easily applied to any network with batch normalization layers, and combined with all other techniques for surprisingly good results.
Image-2-image translation methods coupled with self-regularization loss allows dataset-level refinement. While these methods require additional pipeline to train, it is completely independent and does not require modifications of the main training pipeline.