Size Recommendation: Body Measurement AI on GPU

The Challenge: One in Three Orders Comes Back

An online womenswear brand based in Leeds turns over £12 million annually, but their 31% return rate devours margin. Post-return analysis reveals that 68% of returns cite “wrong size” or “fit not as expected.” The standard size guide — a static table of bust, waist, and hip measurements — fails because customers rarely measure themselves accurately, size charts vary between manufacturers, and the emotional experience of fit cannot be captured in centimetres. Each return costs the brand £8.40 in reverse logistics, restocking, and repackaging, totalling over £600,000 per year in size-related return costs alone.

Third-party virtual fitting room services charge £0.15–£0.50 per recommendation and require customers to upload body images to external servers. For a brand selling intimate apparel alongside casualwear, asking customers to send body photographs to a third-party API creates significant privacy concerns and GDPR compliance risk.

AI Solution: Computer Vision Body Measurement

Body measurement AI uses pose estimation and depth prediction from a single smartphone photograph to estimate key body dimensions — bust, waist, hip, inseam, shoulder width, arm length. Models such as those built on MediaPipe, OpenPose, or custom architectures trained on 3D body scan datasets can predict measurements within 1.5cm accuracy from a front and side photo pair.

The pipeline runs on a dedicated GPU server: the customer takes two photos in their app, the images are sent to the brand’s own infrastructure, the vision model extracts body landmarks and predicts measurements, and a recommendation engine maps those measurements to the best size for each garment based on the brand’s specific fit data. All processing happens on UK-hosted hardware — no body images leave the brand’s control.

GPU Requirements

Pose estimation and body measurement prediction involve multiple model stages: person detection, keypoint estimation, depth prediction, and measurement regression. The full pipeline needs enough VRAM to hold all models simultaneously for low-latency inference.

For a brand processing 5,000 size recommendations per day during peak periods, the RTX 5090 handles the load comfortably. The private AI hosting configuration ensures sensitive body measurement data never leaves controlled infrastructure.

Recommended Stack

• MediaPipe or MMPose for 2D pose estimation and body keypoint detection.
• Depth Anything or MiDaS for monocular depth estimation, converting 2D images to 3D measurements.
• Custom regression model trained on paired data (photos + ground-truth measurements from 3D body scans).
• FastAPI serving the full pipeline with GPU-accelerated inference behind an HTTPS endpoint.
• Optional: an LLM via vLLM to generate natural-language fit advice (“This dress runs slightly small through the hips — we recommend sizing up”).

Brands wanting virtual try-on can extend the pipeline with Stable Diffusion or an image generator to show how a garment drapes on the customer’s predicted body shape.

Cost Analysis

Third-party virtual fitting services at £0.15–£0.50 per recommendation, with 150,000 annual uses, cost £22,500–£75,000 per year. Self-hosting on a dedicated GPU eliminates per-use charges entirely. More importantly, the 40% reduction in size-related returns saves over £240,000 annually in reverse logistics costs.

The secondary benefit is conversion rate improvement. Customers confident in their size are 28% more likely to complete purchase, particularly in categories like dresses and tailored jackets where fit anxiety drives cart abandonment. The brand projects an additional £180,000 in annual revenue from improved conversion.

Getting Started

Begin with a measurement validation study: recruit 200 customers to provide both AI-predicted measurements and professional tailor measurements. Calibrate the model until prediction accuracy falls within 1.5cm for key dimensions. Roll out to a single product category first — dresses typically show the highest size-related return rates — and measure return rate changes over 60 days before expanding site-wide.

GigaGPU provides UK-based dedicated GPU servers configured for vision AI workloads with full data privacy. Add an AI chatbot for personalised styling advice, or deploy additional vision models for outfit recommendation.