Application‑specific AI for your data applications
TOPAS Vision
TOPAS develops and integrates custom AI solutions for companies that need reliable analysis of data, especially image data. We combine methodological expertise with many years of practical experience in robotics, autonomy, perception, and mapping. Our focus is on delivering robust solutions that operate reliably in everyday use and integrate seamlessly into existing processes.
What we offer
End‑to‑end in 3 phases
Process analysis & machine learning concept
We assess goals, KPIs, data, feasibility, and integration.
Proof of concept on real data
Implementation, training, evaluation, and transparent results demonstrate the value.
Integration & Deployment
Integration of interfaces, operations, and monitoring prepare the handover to regular production.
Between these steps, we ensure clarity and predictability: As a starting point, we work with you to define measurable success criteria. We then proceed based on your real-world data, always keeping future integration in mind from day one. This approach prevents a mere ‘AI demo’ and instead delivers a solution that fits your operations both technically and organizationally — with a transparent process and reliable results.
Why TOPAS?
Image‑based machine‑learning core expertise: classification, segmentation, object detection, and complex special cases including post‑processing.
Privacy & control: development and training on our own hardware.
Production‑ready solutions: We don’t just deliver a model, but a complete solution that is ready for immediate use, including all necessary steps for integration.
Edge and real‑time expertise: computing power directly where the data is created. Solutions that work even in resource‑constrained environments.
C++ inference stack: high‑performance execution for demanding applications.
Robustness built on real‑world experience: expertise from real systems such as autonomous machines and robotics.
Real‑world application examples
Small‑object detection & tracking – traffic lights in real time
This example demonstrates how we work when standard models are not sufficient: We adapt not only the weights but also the architecture itself — essential when dealing with special optics, unusual viewing angles, tiny objects, tight latency budgets, or edge hardware.
Traffic‑light detection example: We address a typical small‑object challenge: traffic lights with only 2–3 pixels are reliably detected, their state is determined, and they are tracked consistently across multiple frames.
The architecture is optimized for high resolution with efficient inference:
Fewer layers, finer feature scales — ideal for tiny objects and transferable to other specialized scenarios.
Tracking is also customized (ByteTrack with multi‑label): states and arrow directions are combined consistently.
Result: A robust, real‑time‑capable solution running on Jetson Orin Nano and x86 GPUs — not just a laboratory experiment, but implemented close to production.
Geometry‑based observation fusion of planar structures – semantic maps despite occlusions
Our method merges segmentations from multiple camera views into stable, geometrically accurate maps — even in the presence of traffic, moving objects, and heavy occlusions. This enables entire vehicle fleets to map reliably.
The example shows our roadway reconstruction in public traffic environments. The approach works equally well for other planar structures such as PCBs (Printed Circuit Boards) or general inspection tasks involving occluded areas.
Part of the method is patent‑pending. For customers, this means a field‑tested, customizable system solution that goes far beyond a single ML model and remains robust under real‑world conditions.