TUHH: Projekt 8 - Pickz.ai

Mehr

TUHH > Studium und Lehre > Für Studierende > Fishing for Experience

Pickz.ai

Development of an AI-based System for Reference-Free Biometrics and 3D Eyewear Visualization

Description of the company

pickz ai is a Hamburg-based startup building a web-based AI workflow solution for the digital eyewear consulting process. We connect opticians, eyewear manufacturers, and customers within one data-driven customer journey - both online and offline.

Our mission is to digitize and structure the entire eyewear consulting process - from needs assessment to recommendation and visualization. Instead of isolated tools like Virtual Try-On, we build an integrated, AI-powered workflow that supports opticians, gives manufacturers contextual visibility, and helps customers make confident decisions.

We have already developed several functional prototypes, tested them with 100+ real users, secured LOIs with around 200 opticians, and are working with paying partners. Now, we want to push the technological frontier further - and this is where you come in.

Situation

Current online Virtual Try-On solutions often look impressive at first glance - but they struggle where it truly matters: precision and trust.

In eyewear, millimeter accuracy is essential.

Yet most existing solutions rely on workarounds:

Customers are asked to hold credit cards or other reference objects to their forehead to estimate scale - a process that feels awkward, unreliable, and breaks the user experience.
From a purely mathematical perspective, deriving absolute real-world measurements (in millimeters) from a single 2D image without a physical reference is an ill-posed problem.
Many eyewear products are only available as 2D product images, without depth information or parametric models.

As a result, these solutions prioritize visual appeal over geometric accuracy - offering a stylistic preview rather than medically reliable measurement.

Problem

To unlock a truly intelligent and trustworthy digital eyewear experience, we must address three fundamental technological and user-centric challenges:

Scaling without Reference: Without a physical reference object, it is mathematically challenging to derive absolute real-world dimensions from a 2D image.
2D-to-3D Data Transfer: Creating 3D models manually from 2D templates is expensive, time-consuming, and not scalable.

Solving these challenges is essential to tackle the core problem of eyewear purchasing: information overload. Only if measurements are accurate and visualizations geometrically reliable can an AI-powered Virtual Try-On move beyond simple visualization and truly guide customers - intelligently narrowing hundreds of options down to the few models that genuinely fit their face, needs, and preferences.

Aims of the project

1. Measurement – Reference-Free PD Estimation

Develop an algorithm that estimates pupillary distance (PD) purely from facial cues in a standard camera image - without requiring any external reference object such as a credit card. A promising direction is to use stable biometric features as an implicit scale, for example the iris diameter as a natural biological constant, combined with robust face landmark detection and geometric modeling.

The objective is to turn an everyday selfie into a reliable millimeter-level PD estimate that works in real-world conditions.

If time allows, we will also explore an additional module:

2. Transformation – 2D-to-3D Reconstruction

Build a model that transforms 2D eyewear product images into realistic 3D representations suitable for Virtual Try-On and simulation.

Possible technical approaches:

NeRF-based reconstruction
Depth estimation
Learned parametric shape models

You will work closely with our founding team, gain first-hand insight into building an AI-driven startup, and contribute to a technically challenging problem with real product impact in a scalable system used in the eyewear industry. If you are excited about turning ambitious ideas into working solutions and seeing your code move from prototype to production, we would love to work with you.

Scopes

Depending on your strengths and interests, you may work on:

Researching and implementing face landmark detectors (e.g., MediaPipe, Dlib)
Exploring NeRFs or depth estimation techniques
Developing and validating geometric scaling approaches
Designing evaluation metrics for real-world accuracy and UX quality
Training or fine-tuning 2D-to-3D reconstruction models
If time allows: Building a prototype recommendation system (collaborative or content-based filtering)

This is not a pure research exercise. Your work will directly contribute to a live product with real pilot customers.

Target group (students)

Study Programs: