Executive summary
Context
Van de Velde, a Belgian luxury lingerie group with over a century of craftsmanship and a global presence through brands such as PrimaDonna, Marie Jo, and Andres Sarda, is no stranger to the complexities of fit. In lingerie, small variations in sizing can make or break comfort and confidence, replicating the expertise of an in-store fitting online has long been one of the biggest barriers to e-commerce growth in the sector. Like many fashion and apparel retailers, Van de Velde faced rising return rates, high operational costs, and customer frustration caused by the trial-and-error process of finding the right size remotely.
Goal of the project
The goal of the project was to enable customers to find their perfect bra size measurements at home through a simple, smartphone-based tool. The solution must be privacy-friendly, seamlessly integrated with Van de Velde’s systems, and deliver reliable, cost-effective results comparable to in-store fittings.
Computer Vision solution
To address this, Van de Velde partnered with Superlinear to develop a tool that estimates body measurements using AI and starting from just two smartphone photos and the user’s height. By combining Van de Velde’s deep knowledge of lingerie design with Superlinear’s expertise in AI and computer vision, the companies created a solution that delivers measurement accuracy comparable to professional fittings. Already live on Sardaworld.com, the tool not only reduces costly returns but also enhances customer confidence, positioning Van de Velde at the forefront of innovation in a rapidly evolving retail landscape.
Case study
The challenge
Van de Velde set out to address one of the most complex challenges in online lingerie retail: accurate sizing at scale, accessible to anyone and with the only requirement of taking two smartphone photos. In collaboration, we developed an AI-powered sizing tool that extracts precise body measurements from just two smartphone images combined with a user’s height. The system applies advanced computer vision and statistical modeling to achieve accuracy comparable to in-store fittings. By enabling reliable at-home sizing, the solution reduces return rates, optimizes operational costs, and improves the overall customer experience.
The process
From day one, Superlinear and Van de Velde formed a joint task force to transform an early prototype into a production-ready solution. The AI pipeline was complex, spanning image capture, body segmentation, pose estimation, keypoint detection, and 3D measurement and matching, and each stage needed rigorous testing and tuning to achieve reliable performance.
We worked in sprints, focusing on one stage at a time. Each component got tested under real-world conditions, issues such as segmentation artifacts or pose inconsistencies were identified, and fixes were applied before moving forward.
Through this iterative cycle of test, feedback, and targeted improvement, accuracy and reliability steadily increased. What started as a rough prototype became a robust, production-grade tool, which was made only possible by the close collaboration between Superlinear’s AI experts and Van de Velde’s specialists.
The solution: AI Bra Size Calculator
The end solution consists of a web application that guides customers through the process of taking two simple photos with their phone and entering their height. Behind the scenes, the images are processed via an AI pipeline that estimates key body measurements and confronts them with proprietary 3D mashes.
The entire development followed a sprint-based approach that was used to estimate the best choices to be taken and identify best areas of improvement.
Data collection
Early on, we investigated how user posture and camera conditions affected measurement accuracy. Small changes, such as asking users to stand with arms open or keeping the phone upright, significantly reduced segmentation artifacts and distortions around the bust area. By testing under diverse conditions such as different devices, backgrounds, camera settings, and user environments, we collected realistic data from internal events and validated the robustness of our approach.
Model selection
Choosing the right model was critical. While general solutions for full-body segmentation or keypoint detection already exist, bust-specific segmentation remains unsolved or incomplete. We evaluated multiple models to balance speed and accuracy in a constrained environment. Ultimately, training a custom model gave us the best trade-off: precision, adaptability, and the ability to compute body-specific keypoints not normally available in standard models.
Measurement definitions
Defining what to measure and how was not straightforward. Together with Van de Velde’s experts, we iteratively refined measurement definitions and validation rules. For example, we had to determine which reference points on the body would provide the most accurate results. Should measurements start from the shoulders, extend below the hips, or focus on the distance between the shoulders and the under-bust? These iterations, guided by lingerie fitting expertise, helped us understand which dimensions truly matter for garment fit. In parallel, we ensured that both photos were properly aligned and that segmentation boundaries were precise enough to calculate bust dimensions. By spotting recurring error trends, we could adjust our approach and steadily improve measurements.
Error monitoring and iteration
Even after deployment, learning continued. The production pipeline was instrumented with error codes to quickly flag problematic cases and provide feedback for further improvement. This monitoring loop allowed us to detect, analyze, and fix issues systematically, and turned real usage into an opportunity for continuous improvement.
The results
The solution computes measurements in just a few seconds, making the process fast and accessible for customers. Accuracy is high: 66% of measurements fall within the professional error margin, while the remaining cases are typically off by no more than one size, a strong result for at-home sizing.
In addition, the system generates detailed error logs that allow monitoring and targeted improvements, ensuring that accuracy and reliability will keep increasing over time.
The future
For Van de Velde, the sizing tool is just the beginning. Beyond reducing returns, collecting exact measurements can enable personalized recommendations for styles and cuts based on each customer’s unique body shape, building loyalty and boosting e-commerce conversion.
For the retail sector at large, the implications are significant. Apparel and fashion brands lose billions each year to returns, logistics costs, and unsold inventory linked to sizing uncertainty. AI in online shopping can help retailers cut waste, lower carbon footprints, and create leaner supply chains by aligning production more closely with actual demand. Over time, the technology could evolve into a cross-brand standard for digital fittings, integrated seamlessly into online stores, virtual wardrobes, and even augmented-reality try-on experiences.
Ultimately, this type of innovation pushes the retail sector closer to a future where fit confidence becomes the norm online, bridging the gap between the physical and digital shopping experience.
