Computer Vision Training Data for Industrial Automation

Computer Vision Training Data for Industrial Automation

Close the Sim-to-Real gap and train models that truly handle real world noise.

Trusted by industry leaders & innovators.

Our 3 Step Approach

3D Configuration

Desired objects and vehicles are built into a 3D scene

Ground Truth Data Generation

3D simulation generates RGB, instance and depth maps

Generative AI Augmentation

Photorealistic variation with preserved semantics

Why use Repli5?

0 Days

waiting for logging or annotation of training data

10:1

Repli5-to-Real image ratio for comparable model performance

$2,200

estimated savings per 1000 images vs. market rates

FOUNDATIONAL RESEARCH

State of The Art

A curation selection of technical papers that inform and inspire our approach to training data generation.

04. Domain Randomisation

Westerski & Fong 2025

Synthetic Data for Object Detection with Neural Networks: State-of-the-Art Survey of Domain Randomisation Techniques

View Research

03. Synthetic:Real

Nvidia Research 2025

Optimizing Sim-to-Real Transfer with Asset Randomization and Massive Parallelization

View Research

02. Dataset Generation

Goran Paulin & Marina Ivasic-Kos 2023

Review and Analysis of Synthetic Dataset Generation Methods and Techniques for Application in Computer Vision

View Research

01. Domain Randomisation

NVIDIA Research / CVPR 2018

Training Deep Networks with Synthetic Data: Bridging the Reality Gap by Domain Randomization

View Research

04. Domain Randomisation

Westerski & Fong 2025

Synthetic Data for Object Detection with Neural Networks: State-of-the-Art Survey of Domain Randomisation Techniques

View Research

03. Synthetic:Real

Nvidia Research 2025

Optimizing Sim-to-Real Transfer with Asset Randomization and Massive Parallelization

View Research

02. Dataset Generation

Goran Paulin & Marina Ivasic-Kos 2023

Review and Analysis of Synthetic Dataset Generation Methods and Techniques for Application in Computer Vision

View Research

01. Domain Randomisation

NVIDIA Research / CVPR 2018

Training Deep Networks with Synthetic Data: Bridging the Reality Gap by Domain Randomization

View Research

Pixel-perfect alignment.

Our annotations are not the added to the image. They are the DNA of the scene. Generate bounding boxes or semantic labels natively aligned to the geometry.

WEATHER: Clear

Environment:Town_10

REVEAL_GROUND_TRUTH

REVEAL_DAYLIGHT_VARIATION

Environment:Scandinavian_Forest

Time_of_Day:Evening

Long tail training data

A single semantic frame can spawn infinite visual variations. Change backgrounds, lighting, reflections, add scratches, residual lubricant or production mill marks.

Domain Randomisation

Domain Randomisation

Generate the impossible to break model biases. Decouple object features from background noise to improve model detection in the field.

Hello / Good morning

Bonjour

WEATHER_VARIATION

Environment:NORDIC_COASTAL

REVEAL_SCENE

> READY_FOR_CONNECTION_ESTABLISHMENT...
> READY_FOR_CONNECTION_ESTABLISHMENT...

Generate your next training dataset.

Curate computer vision training data in days, not months.
[STATUS:PENDING_CONTACT]

Kungsgatan 10a
419 19 Gothenburg
Sweden

Kungsgatan 10a
419 19 Gothenburg
Sweden

Kungsgatan 10a
419 19 Gothenburg
Sweden