The $100 Billion Problem Hiding in Plain Sight

Every year, the fashion industry loses billions of dollars to a problem that seems almost too basic to exist: garments that don't fit.

Return rates for apparel e-commerce can reach 30% or higher, and for made-to-measure fashion, the numbers are even more staggering—often climbing to 80%. The culprit is almost always the same: inaccurate customer measurements.

A customer pulls out a tape measure, wraps it around their body, guesses at the right tension, rounds up or down, and submits numbers that are, more often than not, wrong. The garment arrives. It doesn't fit. The customer is disappointed. The brand eats the cost of return shipping, restocking, and lost trust.

This is the problem SartorFit set out to solve.

Our solution is an AI-powered body measurement system that uses nothing more than a standard smartphone camera. No tape measure. No guesswork. No training required. The customer simply stands in front of their phone, follows a few simple instructions, and the AI captures precise measurements in under two minutes.

But developing an AI system that can accurately measure the human body from a photograph is not a trivial task. It requires vast amounts of high-quality training data—thousands of images of real human bodies, paired with precise ground-truth measurements. Without that data, the AI cannot learn to generalize across the incredible diversity of human body shapes, sizes, and proportions.

This is where Amazon's BodyM dataset enters the picture.

What Is the BodyM Dataset?

BodyM is the first large-scale public body measurement dataset, released by Amazon's research team to advance the field of human body measurement from images. It consists of 8,978 frontal and lateral silhouette images captured from 2,505 real human subjects, each paired with height, weight, gender, and 14 precise body measurements.

The measurements include:

• Ankle girth

• Arm length

• Bicep girth

• Calf girth

• Chest girth

• Forearm girth

• Height

• Hip girth

• Leg length

• Shoulder breadth

• Shoulder-to-crotch length

• Thigh girth

• Waist girth

• Wrist girth

Each subject in the dataset is represented by two black-and-white silhouette images—one taken from the front and one from the side—with the full body in view. These silhouettes strip away the visual noise of clothing, lighting, and background, allowing the AI to focus on the essential shape of the body.

The dataset is available on the AWS Open Data Registry and is licensed under a Creative Commons Attribution-NonCommercial 4.0 International Public License. This means researchers and developers can access and use the data for non-commercial purposes, advancing the state of the art in body measurement estimation.

For SartorFit, the release of BodyM was a turning point.

Why BodyM Matters for Fashion AI

Prior to BodyM, most body measurement datasets were either small, private, or focused on synthetic data. Researchers and developers had limited access to real-world measurements from diverse human subjects. This made it difficult to build AI systems that could generalize across the full spectrum of body shapes.

BodyM changes that. It provides a large, diverse, and publicly available dataset that enables researchers to train and evaluate body measurement models with unprecedented rigor. The dataset includes subjects of various shapes and sizes, with some photographed multiple times wearing different clothing to test robustness.

The research paper accompanying BodyM, "Human Body Measurement Estimation with Adversarial Augmentation," describes a novel approach to training body measurement models. The authors present a Body Measurement network (BMnet) that estimates 3D anthropomorphic measurements from silhouette images, height, and weight.

The key innovation is the Adversarial Body Simulator (ABS) —a fully differentiable module that generates challenging body shapes to expose gaps in the training data. The ABS is based on the SMPL (Skinned Multi-Person Linear) body model and aims to maximize the BMnet's measurement prediction error. By training BMnet with adversarial examples, the model becomes more robust and accurate, particularly for under-represented body shapes.

The results are impressive. Training BMnet with ABS improves measurement prediction accuracy on real bodies by up to 10% compared to no augmentation or random body shape sampling. The method significantly outperforms state-of-the-art measurement estimation methods by as much as 3x.

From Dataset to Deployment: SartorFit's Journey

SartorFit's body measurement AI, known as SartorGuide AI, is built on the foundations laid by the BodyM dataset and the BMnet architecture. Our team has worked tirelessly to adapt and extend these technologies for the specific needs of the fashion industry.

The Training Process

The first step in building SartorGuide AI was training our models on the BodyM dataset. The 8,978 silhouette images and 14 corresponding measurements per subject provided the ground-truth data needed to teach the AI how to infer body measurements from visual input.

We used a convolutional neural network architecture similar to BMnet, with additional customizations to improve accuracy for the measurements most critical to fashion fit—chest, waist, hip, shoulder breadth, and inseam. The network was trained to take a frontal and side silhouette image as input, along with the subject's height and weight, and output the 14 body measurements.

Adversarial Training for Robustness

One of the most valuable aspects of the BodyM research is the Adversarial Body Simulator. The ABS generates synthetic body shapes that are challenging for the measurement network, forcing it to improve its accuracy across the full spectrum of human bodies.

We have implemented a similar adversarial training regime at SartorFit. By continuously generating challenging body shapes and retraining our models, we have been able to achieve measurement accuracy that approaches that of a professional tailor.

Beyond Silhouettes: Real-World Images

While the BodyM dataset uses clean silhouette images, real-world users will take photos of themselves in everyday clothing, with varying lighting, backgrounds, and poses. To bridge this gap, we have augmented our training data with additional datasets and synthetic data that simulate real-world conditions.

Our models have been trained to handle diverse clothing types, lighting conditions, and camera angles. The result is a system that works reliably in the hands of everyday consumers, not just in controlled laboratory conditions.

The Results: Accuracy That Rivals Professional Measurement

The performance of SartorGuide AI speaks for itself. Our internal testing has shown that the system achieves measurement accuracy within 1 to 2 centimeters of professional manual measurement for most key body dimensions—chest, waist, hip, and inseam.

In some cases, the AI actually outperforms manual measurement. Professional tailors, after all, are human. They make mistakes. They have off days. They rush. The AI never does.

To put this in perspective, the original BMnet paper reports Mean Absolute Error (MAE) as low as 0.8 mm for ankle girth, 1.3 mm for forearm girth, and 1.7 mm for bicep girth and shoulder-to-crotch length. Overall measurement errors are reduced by up to 70% compared to previous methods.

While SartorGuide AI's performance on real-world images is not quite at the level of the paper's controlled experiments, we are making rapid progress. Our latest internal tests show that we are approaching the accuracy needed to deliver reliable made-to-measure garments with minimal human intervention.

A Glimpse Behind the Curtain: Internal Testing Demo

We have documented our internal testing process in a video demonstration, available on the SartorFit YouTube channel.

Watch the Internal Testing Demo on YouTube

The demo shows: https://www.youtube.com/watch?v=pBQK0RtctfA

• Data Collection: How we capture silhouette images and ground-truth measurements from test subjects.

• Model Inference: The AI processing a silhouette image and outputting the 14 body measurements.

• Comparison: Side-by-side comparison of AI measurements and manual measurements taken by a professional tailor.

• Error Analysis: Visualization of measurement errors and identification of areas for improvement.

• Real-World Testing: The AI processing a photo of a person in everyday clothing, rather than a clean silhouette.

The video provides a transparent look at where we are in our development journey—the successes, the challenges, and the path forward.

The Quality Trajectory: Where We're Headed

SartorGuide AI is already good. But it is about to get much better.

More Data, Better Models

We are continuously expanding our training dataset with additional real-world images and measurements. The more data we collect, the more accurate our models become. We are also exploring the use of synthetic data—generated by adversarial networks—to fill gaps in our training distribution.

Fine-Tuning for Fashion

Not all measurements are equally important for fashion fit. Chest, waist, and hip measurements are critical for most garments, while ankle and wrist girth are less so. We are fine-tuning our models to prioritize the measurements that matter most for garment fit and comfort.

Integration with 3D Body Modeling

The ultimate goal is to create a complete 3D body model from a few smartphone photos. This would allow customers to see exactly how a garment will fit and look on their own body before they buy it. We are actively working on integrating SartorGuide AI with 3D body modeling technologies.

Reducing Returns, Building Trust

The end goal is simple: reduce made-to-measure return rates from 80% to under 15%. Every percentage point reduction in returns translates into millions of dollars in savings for fashion brands, not to mention the environmental benefit of reducing waste.

The Broader Impact: A New Era for Fashion

SartorGuide AI is not just a technical achievement. It is a fundamental shift in how fashion works.

For Consumers

No more measuring tape. No more guesswork. No more returns. Customers can order made-to-measure garments with confidence, knowing that the garment will fit perfectly the first time.

For Fashion Brands

Made-to-measure becomes scalable. Brands can offer custom fit to every customer, not just a select few. Return rates plummet. Customer satisfaction soars. Profit margins expand.

For the Environment

Fewer returns means less shipping, less packaging, and less waste. Made-to-measure production is inherently more sustainable than mass production, as garments are made to order rather than in excess.

For the Industry

Body measurement AI levels the playing field. Small brands and independent designers can offer the same fit accuracy as luxury houses. The technology democratizes made-to-measure fashion.

Conclusion: The Perfect Fit Is Finally Possible

Made-to-measure fashion has been broken for too long. High return rates. Frustrated customers. Struggling brands. The promise of custom fit has never matched the reality.

Amazon's BodyM dataset and the BMnet architecture have provided the foundation for a new generation of body measurement AI. SartorFit is building on that foundation to create a system that is accurate, user-friendly, and scalable.

SartorGuide AI is already achieving measurement accuracy within 1 to 2 centimeters of professional manual measurement. With continued training, fine-tuning, and integration with 3D body modeling, we are on track to reduce made-to-measure return rates from 80% to under 15%.

The perfect fit is finally possible.

One measurement at a time.

Links Summary

ResourceURLSartorFithttps://sartorfit.comLolaGuildhttps://lolaguild.comSartorGuide AIhttps://ai.sartorfit.comBodyM Datasethttps://registry.opendata.aws/bodym/Internal Testing DemoWatch on YouTube

References

1. Ruiz, N., Bellver, M., Bolkart, T., Arora, A., Lin, M. C., Romero, J., & Bala, R. (2022). Human Body Measurement Estimation with Adversarial Augmentation. Amazon Science.

2. BodyM Dataset. (n.d.). AWS Open Data Registry. Retrieved from https://registry.opendata.aws/bodym/

3. BodyM Dataset. (n.d.). DagsHub. Retrieved from https://dagshub.com