How reverse.fashion Scales Textile Sorting AI

How reverse.fashion Scales Textile-Sorting AI with Voxel51

reverse.fashion builds visual AI systems that automate textile sorting for reuse and recycling. Its five-person machine learning team works with millions of ultra-high-resolution garment images, far more than any team could hand-label. reverse.fashion uses Voxel51 as a shared visual layer to find the samples worth labeling, keep labels aligned with the real sorting task, and avoid hundreds of hours of infrastructure work.

Category	Details
Industry	Textile sorting for reuse and recycling
Region	Europe
ML team	Five engineers, plus a product owner with textile expertise
Data	Millions of ultra-high-resolution, two-sided garment images
Models	Defect detection, fabric and product classification, style recognition
Journey	FiftyOne open source → upgraded to Voxel51 on an enterprise plan
Stack	Cloud object storage, VS Code and cloud notebooks, custom labeling backends
Key results	Hundreds of engineering hours saved, dataset reviews without notebook handoffs

reverse.fashion and Voxel51 at a glance.

reverse.fashion: visual AI for textile sorting

reverse.fashion is building computer vision systems to automate textile sorting. Across Europe, enormous volumes of used clothing flow into collection systems that still depend heavily on manual sorting.

Its systems use two-sided, ultra-high-resolution cameras to scan multiple garments moving at several meters per second. From those images, reverse.fashion models assess garment quality and detect defects, identify fabric and product types, and recognize visual styles that help determine where an item belongs next.

For reusable and vintage garments, that can mean identifying characteristics that make an item valuable to high-end secondhand markets. For recycling, it can mean understanding fabric type and other properties needed to determine the appropriate downstream process.

reverse.fashion's line.sort scans garments at high speed with a double-sided, ultra-high-resolution camera.

The challenge: how to select the most valuable samples to label

For reverse.fashion, collecting visual data isn't the hard part. The company has more imagery than its 5-person machine learning team could ever reasonably hand-label.

The harder problem is deciding what should be labeled, how it should be labeled, and whether those labels actually reflect the real-world sorting task.

Textiles don't come with a clean, universal ontology. Categories can overlap. Garment condition can be subjective. Definitions of defects and other attributes can vary depending on the downstream application.

That makes visual inspection critical. A model can perform well against a metric while still solving the wrong practical problem.

“You can fulfill the metric, but you can fail on the application.” — Dr. Karsten Pufahl, CTO, reverse.fashion

For reverse.fashion, dataset quality therefore comes down to data curation, an iterative loop: find the right examples, inspect them visually, refine the labeling strategy, evaluate model behavior, and confirm that what the model learned matches what the application actually needs.

Annotating style, not just defects

reverse.fashion's models detect and localize garment defects such as stains and thinning, with per-defect confidence scores.

A garment isn't simply “good” or “bad.” reverse.fashion's models may need to distinguish between different stains, defects, materials, construction types, product categories, and style attributes, sometimes at an extremely fine level of detail.

The team has experimented with distinctions as specific as different kinds of stains and even human versus animal hair. That level of granularity makes visual dataset inspection essential: the team continually has to determine which distinctions are meaningful enough to teach a model, and which aren't.

reverse.fashion chose Voxel51 to focus on building models, not infrastructure

reverse.fashion began using Voxel51 early in its development, before data management had become a major bottleneck.

When the team first adopted Voxel51, its datasets were smaller, image resolution was lower, and local development was still manageable. But reverse.fashion knew its data needs would become more complex as the company moved from research toward production.

Different computer vision problems also meant working across diverse datasets and annotation formats, including bounding boxes, segmentations, keypoints, and custom ontologies. Building internal tooling to normalize, query, and visualize all of those datasets would have pulled engineering effort away from the company's actual differentiator.

“We wanted to focus on building models and not building infrastructure.” — Dr. Karsten Pufahl, CTO, reverse.fashion

reverse.fashion started with the open source version of FiftyOne. As image resolution increased, hardware scaled, defect detection became a larger part of the workload, and more people needed to work with the same datasets, the team moved to Voxel51.

Voxel51 provided the synchronized, shared access the growing team needed: multiple people could inspect the same datasets, access one another's labels, and discuss what they were seeing without packaging the work into one-off notebooks or exports.

How reverse.fashion uses Voxel51

Voxel51 now serves as a shared visual layer across several parts of reverse.fashion's ML workflow.

The team works against imagery stored in cloud object storage, pulls new data into datasets through custom tooling, integrates external labeling workflows, and develops through both local environments and cloud notebooks.

Voxel51 also gives engineers a visual interface for inspecting model outputs and embeddings. The team uses dimensionality reduction and embedding exploration to identify interesting regions of a dataset while staying connected to the underlying images.

For reverse.fashion, that combination matters: engineers can move between code, queries, model outputs, and the actual visual examples without creating a separate inspection workflow for every task.

A hub for a growing MLOps stack

reverse.fashion connects Voxel51 to cloud object storage and has built custom tooling to bring new data into its datasets. The team has also connected labeling workflows and built custom label backends for cases where its textile ontologies required more control over valid label combinations.

The development workflow remains flexible. Some engineers prefer working locally in Visual Studio Code, while others use cloud notebooks. Voxel51 supports both while keeping the underlying datasets accessible through a shared environment.

reverse.fashion's computer vision workflow: five steps that close the loop between the catalog, the labels, and what the model actually learned.

reverse.fashion models also identify labels, buttons, and other garment attributes that feed reverse.fashion's textile taxonomy.

Bringing domain experts into the loop

Not every annotation problem can be solved by sending more images to a labeling vendor.

Some examples require genuine textile expertise.

reverse.fashion's product owner comes from a textile background and can use Voxel51 directly to inspect data and contribute high-quality labels when the ML team encounters difficult or subjective examples.

“Sometimes it needs really high-quality labeling… and then it's good that you can do it tomorrow.” — Dr. Karsten Pufahl, CTO, reverse.fashion

Results: a five-person ML team finds the signal in millions of textile images

Before Voxel51, reviewing another developer's dataset could mean preparing Jupyter notebooks, generating contact sheets or video examples, adjusting loops to surface the right samples, exporting thousands of images into folders, and manually scrolling through the results.

High-resolution imagery made the process even harder. Large datasets could exceed what was practical to render in a notebook, and simply preparing the right visual examples for another person could consume hours.

Building the broader infrastructure needed to query, visualize, zoom into, and review this data internally would require hundreds of engineering hours.

“Before, we were building contact sheets, rendering images into folders, and trying to review thousands of samples in notebooks. Now, we log in and look at the data.” — Dr. Karsten Pufahl, CTO, reverse.fashion

With Voxel51, that collaboration takes minutes: log in, run a query when needed, and look at the data.

That shared visibility changes more than review speed. Developers can inspect one another's datasets and labels without preparing custom handoffs. Textile experts can review difficult examples directly. And the ML team can use visual inspection to make sure its labels, models, and evaluation criteria remain aligned with the actual sorting problem.

Shared visibility: reverse.fashion’s developers can access the same datasets and labels without preparing custom notebook or folder-based handoffs.
Better label quality: Visual review helps the reverse.fashion team verify that annotations and model behavior match the real-world textile problem.
Domain expertise on demand: reverse.fashion’s textile specialists can directly inspect difficult examples and contribute feedback alongside the ML team.

“If you have a lot of data and you’re more than one person, I would say Voxel51. I don’t know any better solution for this.” — Dr. Karsten Pufahl, CTO, reverse.fashion

FAQ

reverse.fashion uses Voxel51 as a shared visual layer for dataset exploration, embedding analysis, labeling integration, and model evaluation across its five-person ML team.

The team moved to Voxel51 when rising image resolution, a growing defect-detection workload, and more people needing the same datasets required synchronized, shared access.

reverse.fashion runs an iterative curation loop: find the right examples, inspect them visually, refine the labeling strategy, evaluate model behavior, and confirm the model matches the application.

Yes. reverse.fashion's product owner, who comes from a textile background, uses Voxel51 directly to inspect data and contribute labels on difficult or subjective examples.

How reverse.fashion Scales Textile-Sorting AI with Voxel51

Talk to a computer vision expert

How reverse.fashion Scales Textile-Sorting AI with Voxel51

reverse.fashion: visual AI for textile sorting