inTens-Vision is an end-to-end platform focused exclusively on automated part inspection and dimensional accuracy. We train neural networks 100 % on synthetic CAD data in order to capture parts via 6D pose estimation — purely for exact spatial localisation. The actual measurement of dimensional accuracy (tolerances, distances, edge dimensions) is then performed by classic, subpixel-accurate edge tracing – executed highly efficiently on an NVIDIA GPU right inside the control cabinet.
Hybrid approach: AI locates the part, classic mathematics measures it.
System architecture
End-to-end AI vision platform across three core domains
Central management & training in the web UI, edge inference and metrology on the Jetson, data storage and traceability in a modern NoSQL structure – designed as one continuous loop, not as an isolated solution.
01 Web UI · Cloud
Central management, CAD database & synthetic training
Everything starts with the design data – not with the production line. The inspection task is created virtually while the factory is still being built.
STEP / IGES
ONNX / TensorRT
React Flow
Domain randomization
Central CAD database
Central management of all 3D CAD models (STEP/IGES) – logically linked to the neural networks trained on them, their weights (ONNX/TensorRT) and the matching inspection configurations.
Virtual annotation (no-code)
Users upload the CAD model and mark the features, edges, holes and distances to be inspected by drag and drop – directly on the 3D model or on rendered ideal views.
Automated synthetic pipeline
An integrated renderer automatically generates thousands of photorealistic training images from the CAD model. Physically photographing parts is no longer necessary at all.
Domain randomization (sim-to-real)
While rendering, the software automatically varies textures, metallic reflections, image noise, shadows and changing ambient light scenarios. The CNN becomes robust for daily shop-floor conditions without a single real photo.
Visual pipeline editor
Using an intuitive flow chart, users define the logical inspection sequence: trigger → pose estimation → measure feature A → tolerance check → inspection report.
02 Edge · Control cabinet
Edge inference & hybrid metrology
On the line, two methods work hand in hand: the AI understands the position of the part, classic image processing delivers the incorruptible measurement.
NVIDIA GPU
3 × 2 MP global shutter
Telecentric optics
Subpixel
Hardware basis
A compact NVIDIA GPU for high-performance edge inference directly inside the control cabinet – connected to up to three 2-megapixel cameras: global shutter against motion blur, telecentric lenses for edges free of perspective and distortion.
6D pose estimation (CNN)
The CNN determines the exact three-dimensional position and orientation of the part (X, Y, Z + rotation) within milliseconds. Thanks to the intrinsic camera calibration stored at the factory, the pose is translated directly into metric data – the error-prone alignment relative to the conveyor belt is no longer needed.
Hybrid edge tracing
The inspection features defined in the web UI are projected onto the real edge image with mathematical precision, based on the live 6D pose. A rule-based edge detection algorithm traces the contours with subpixel accuracy and delivers the final, tamper-proof millimetre measurements.
03 OT / IT infrastructure
Deployment, communication & data storage
What works on one line runs on all of them with a single click. And every measurement is retained as a structured data record.
OTA update
OPC UA / PROFINET / MQTT
NoSQL
Cloud-native OTA deployment
Fully automated compilation of the pipeline including model optimisation via NVIDIA TensorRT – containers and models are then pushed straight to the edge devices in the factories via OTA update.
Line PLC integration
Native OT interfaces (OPC UA, PROFINET, MQTT) on the Jetson transmit measurement results – OK/NOK and tolerance values – to the higher-level machine control within milliseconds.
NoSQL database & traceability
Structured storage of all measurement reports (timestamp, part number, metric values), linked to the compressed image crops of the measured contours. Complete traceability and statistical process control (SPC) for early detection of tool wear.
We take the sim-to-real problem seriously
The leap from a perfect CAD rendering to a real image – scratches in the metal, minimal differences in reflection – is the biggest mathematical hurdle for a CNN. We address it with rigorous domain randomization: shadows, noise and textures are varied extremely during rendering so the network does not fail in a real factory.
Measuring stays mathematics, not an AI estimate
Industry rightly does not yet trust "pure AI" for measurement. That is why our approach is deliberately hybrid: the AI handles 6D pose detection, the actual measurement comes from classic subpixel edge tracing. Comprehensible, reproducible, auditable.
Pipeline & data flow
From the digital twin to edge inference close to production.
01STEP / IGES
CAD model
Import of native design data.
02Domain randomization
Synthetic rendering
AI image generation & variance.
03ONNX export
CNN training
Model optimisation in the cloud.
04FP16 / INT8
TensorRT build
Compilation for NVIDIA hardware.
05Fleet deployment
OTA push
Secure distribution to the edge.
066D pose estimation
Jetson inference
Real-time pose detection on site.
07Subpixel accuracy
Edge tracing
Precision measurement algorithms.
08OK/NOK + report
NoSQL database
Centralised quality records.
Comparison matrix
inTens-Vision in a technology comparison
Because of their sluggish legacy architectures, established machine vision vendors leave open exactly the weak points we address. Nine criteria in direct comparison.
Criterion
inTens-VisionPlatform
Classic vision libraries
Established smart cameras
Proprietary edge systems
Data basis (training)
100 % synthetic (CAD)
Ready to go before the first real part exists – thanks to automated rendering including domain randomization.
limited
Support CAD import, but generating synthetic data requires complex manual scripts and render pipelines.
not available
Rely almost exclusively on real camera images; CAD import for AI training is severely limited.
not available
Requires physical image training directly on the production line; assumes real good and bad parts are available on site.
Part localisation
6D pose estimation
Locates parts with mathematical precision in 3D space (X, Y, Z + rotation) – using nothing but an inexpensive 2 MP camera.
limited
Offer 3D surface matching, but require an extremely high programming effort and calibration expertise.
not available
Primarily classic 2D alignment. True 3D pose detection only via expensive special hardware.
not available
Primarily 2D alignment. Three-dimensional detection only via integrated profile or triangulation lasers.
On-site calibration
Zero calibration (plug & play)
A one-time intrinsic calibration stored at the factory is enough. No manual calibration relative to the conveyor plane required.
not available
High on-site alignment effort; require manually written calibration scripts and precise targets.
not available
Elaborate on-site alignment using calibration plates is mandatory in order to convert pixels into millimetres.
limited
Proprietary but rigid on-site calibration; sensitive to later mechanical vibrations of the equipment.
Hardware / inference
NVIDIA Jetson
Extremely compact, cost-efficient and dedicated to AI multi-stream pipelines inside the control cabinet.
limited
Require an industrial PC (IPC) with a large, expensive and power-hungry NVIDIA desktop GPU in the cabinet.
limited
Proprietary smart cameras (In-Sight) or dedicated Windows vision PCs with high IT maintenance effort.
not available
Proprietary, completely closed controller hardware with extremely high acquisition costs (premium pricing).
Multi-camera capability
Up to 3 streams (2 MP)
Simultaneous, parallel processing of several viewing angles on a single, inexpensive Jetson.
comparable
Scales without limits – but requires correspondingly expensive frame grabber cards and IPC hardware.
not available
Only 1 camera on autonomous smart cameras; multiple streams require expensive PC-based licences.
comparable
Up to 8 cameras can be connected to a single controller.
Deployment / fleet
Over-the-air (OTA)
A central web UI pushes new pipelines and optimised models to all Jetsons remotely, in one click.
not available
Manual copying of code, scripts and binaries onto every single IPC on site.
limited
First approaches via proprietary platforms (OneVision), but functionally very limited and tied to the hardware.
not available
No central fleet management whatsoever; updates only locally at the machine.
Configuration (UI)
Web UI with drag & drop
An intuitive React flow chart for visual logic pipelines – usable without any programming knowledge.
not available
HDevelop IDE; a highly complex development environment that requires deep programming skills.
limited
Spreadsheet logic (Excel-style) or VisionPro Designer – a classic Windows desktop application.
limited
Proprietary desktop software or rigid clicking through a handheld console on the controller.
Data storage & traceability
Modern NoSQL database
Central, highly flexible storage of structured measurement reports – directly linked to compressed image crops.
not available
No native database; the integrator has to build everything themselves via SQL or file paths.
not available
Local FTP storage in Windows directories or expensive, proprietary add-on databases.
not available
Proprietary storage on SD cards or network drives; no flexible NoSQL structure for big data analytics.
Active learning (closed loop)
Automated feedback loop
Borderline cases at the tolerance limits are flagged and flow back into the web UI automatically for model fine-tuning.
not available
Missing entirely: data collection and re-training have to be set up manually as a separate software project.
not available
Missing entirely: data flows end at the camera; there is no automated return channel for AI optimisation.
not available
Missing entirely: once trained, the system stays rigidly at the state of its initial commissioning.
Data basis (training)
inTens-Vision
100 % synthetic (CAD)
Ready to go before the first real part exists – thanks to automated rendering including domain randomization.
Classic vision libraries · limited
Support CAD import, but generating synthetic data requires complex manual scripts and render pipelines.
Established smart cameras · not available
Rely almost exclusively on real camera images; CAD import for AI training is severely limited.
Proprietary edge systems · not available
Requires physical image training directly on the production line; assumes real good and bad parts are available on site.
Part localisation
inTens-Vision
6D pose estimation
Locates parts with mathematical precision in 3D space (X, Y, Z + rotation) – using nothing but an inexpensive 2 MP camera.
Classic vision libraries · limited
Offer 3D surface matching, but require an extremely high programming effort and calibration expertise.
Established smart cameras · not available
Primarily classic 2D alignment. True 3D pose detection only via expensive special hardware.
Proprietary edge systems · not available
Primarily 2D alignment. Three-dimensional detection only via integrated profile or triangulation lasers.
On-site calibration
inTens-Vision
Zero calibration (plug & play)
A one-time intrinsic calibration stored at the factory is enough. No manual calibration relative to the conveyor plane required.
Classic vision libraries · not available
High on-site alignment effort; require manually written calibration scripts and precise targets.
Established smart cameras · not available
Elaborate on-site alignment using calibration plates is mandatory in order to convert pixels into millimetres.
Proprietary edge systems · limited
Proprietary but rigid on-site calibration; sensitive to later mechanical vibrations of the equipment.
Hardware / inference
inTens-Vision
NVIDIA Jetson
Extremely compact, cost-efficient and dedicated to AI multi-stream pipelines inside the control cabinet.
Classic vision libraries · limited
Require an industrial PC (IPC) with a large, expensive and power-hungry NVIDIA desktop GPU in the cabinet.
Established smart cameras · limited
Proprietary smart cameras (In-Sight) or dedicated Windows vision PCs with high IT maintenance effort.
Proprietary edge systems · not available
Proprietary, completely closed controller hardware with extremely high acquisition costs (premium pricing).
Multi-camera capability
inTens-Vision
Up to 3 streams (2 MP)
Simultaneous, parallel processing of several viewing angles on a single, inexpensive Jetson.
Classic vision libraries · comparable
Scales without limits – but requires correspondingly expensive frame grabber cards and IPC hardware.
Established smart cameras · not available
Only 1 camera on autonomous smart cameras; multiple streams require expensive PC-based licences.
Proprietary edge systems · comparable
Up to 8 cameras can be connected to a single controller.
Deployment / fleet
inTens-Vision
Over-the-air (OTA)
A central web UI pushes new pipelines and optimised models to all Jetsons remotely, in one click.
Classic vision libraries · not available
Manual copying of code, scripts and binaries onto every single IPC on site.
Established smart cameras · limited
First approaches via proprietary platforms (OneVision), but functionally very limited and tied to the hardware.
Proprietary edge systems · not available
No central fleet management whatsoever; updates only locally at the machine.
Configuration (UI)
inTens-Vision
Web UI with drag & drop
An intuitive React flow chart for visual logic pipelines – usable without any programming knowledge.
Classic vision libraries · not available
HDevelop IDE; a highly complex development environment that requires deep programming skills.
Established smart cameras · limited
Spreadsheet logic (Excel-style) or VisionPro Designer – a classic Windows desktop application.
Proprietary edge systems · limited
Proprietary desktop software or rigid clicking through a handheld console on the controller.
Data storage & traceability
inTens-Vision
Modern NoSQL database
Central, highly flexible storage of structured measurement reports – directly linked to compressed image crops.
Classic vision libraries · not available
No native database; the integrator has to build everything themselves via SQL or file paths.
Established smart cameras · not available
Local FTP storage in Windows directories or expensive, proprietary add-on databases.
Proprietary edge systems · not available
Proprietary storage on SD cards or network drives; no flexible NoSQL structure for big data analytics.
Active learning (closed loop)
inTens-Vision
Automated feedback loop
Borderline cases at the tolerance limits are flagged and flow back into the web UI automatically for model fine-tuning.
Classic vision libraries · not available
Missing entirely: data collection and re-training have to be set up manually as a separate software project.
Established smart cameras · not available
Missing entirely: data flows end at the camera; there is no automated return channel for AI optimisation.
Proprietary edge systems · not available
Missing entirely: once trained, the system stays rigidly at the state of its initial commissioning.
Market gaps
Five gaps we close in practice
Every one of these gaps costs manufacturers time, money or data quality today. On the left the market status quo, on the right our technical answer.
01
The time gap
Zero time to production thanks to CAD models
The challenge
Production lines often have to stand still, or prototype parts have to be built laboriously, just to capture hundreds of images for AI training.
Our solution
Because we train purely on synthetic CAD data, our software is ready before the first real part comes off the machine. Customers set up the inspection pipeline virtually while the factory is still being built – simultaneous engineering instead of waiting.
0 days
downtime for data collection
02
The complexity gap
True 3D pose detection with 2D cameras
The challenge
When complex parts tilt or rotate in space, classic 2D measurement windows can fail. As a rule this is solved with extremely expensive 3D hardware (laser, structured light).
Our solution
Via 6D pose estimation our system understands how the part is tilted in space from the image of a 2-megapixel camera alone. Virtually annotated edges are projected back onto the real 3D pose with mathematical precision: 3D intelligence at the price of 2D hardware.
2 MP
camera instead of a 3D laser
03
The scaling gap
OTA & fleet management instead of USB sticks
The challenge
Managing machine vision systems in factories feels like IT from 2005: updates are installed locally via USB stick or through rigid network drives.
Our solution
Our OTA architecture on top of the web UI breaks these silos open. A single quality engineer at headquarters creates an optimised pipeline and rolls it out to 50 Jetson systems in factories worldwide with one click.
1 click
for the entire fleet
04
The hardware cost trap
Embedded AI with Jetson instead of premium controllers
The challenge
Highly specialised controllers quickly cost €20,000 and more. Industrial PCs with expensive NVIDIA graphics cards are often required in the control cabinet.
Our solution
Our standard hardware is an extremely cost-efficient powerhouse with a module price below €1,000. The fact that it processes up to three 2 MP global shutter streams in parallel cuts hardware costs massively – and makes the system attractive for mid-sized companies too.
< €1,000
module price instead of a €20,000 controller
05
The data treasure gap
Modern NoSQL structure instead of Windows folders
The challenge
Established systems often discard image data or dump it unstructured into Windows folders. Searching millions of measurements for patterns – "show me every image where dimension X deviated by more than 0.2 mm" – becomes impossible.
Our solution
By natively storing measurement reports together with the exact image crops in a NoSQL database (MongoDB, CouchDB), we enable our customers to do real predictive maintenance and gapless traceability.
100 %
analysable measurement history
Core selling points
Five USPs that make the difference
Why quality managers, maintenance teams and purchasing can all agree on this platform.
01
Zero time to production
Complete independence from real image data
No production downtime, no photographing prototypes to collect training data. The system is 100 % CAD-based and ready before the first real part comes off the line.
02
3D precision at a 2D price
AI-based 6D pose instead of expensive 3D sensors
The 6D pose estimation dynamically computes positional changes and tilts in space. That eliminates the need for expensive and delicate 3D profile laser sensors or structured light projectors.
03
Zero calibration & plug-and-play
MTTR close to zero when service is needed
Combining factory-side intrinsic calibration with a live 6D pose removes the rigid alignment of the camera to the conveyor plane. In a service case the camera is simply swapped, without any local calibration effort.
04
Enterprise fleet management
Central control instead of isolated systems
No more servicing individual industrial PCs by USB stick. New inspection pipelines and optimised models are rolled out worldwide to every line via OTA update from a central web UI, in one click.
05
Maximum acceptance through a hybrid approach
AI locates, classic mathematics measures
A perfect division of labour: the AI handles localisation and understanding of the complex part, while classic subpixel mathematics (edge tracing) computes the dimensions incorruptibly. That creates immediate trust with quality managers.
Contact
Clarify feasibility – without a single physical part
Because we train purely on CAD data, we can assess your inspection task before any prototype exists. Describe your use case – we reply within one business day.
01
CAD model & inspection features
You send us a STEP/IGES model together with the required dimensions and tolerances.
02
Synthetic training
We render the dataset, train the network and annotate the measurement features virtually.
03
Feasibility report
You receive a reliable statement on achievable accuracy, cycle time and hardware requirements.