Executive summary — the one-minute version
|
🏢 One platform, every camera form The same OpenNCC platform shipped as a 38 × 38 mm core board, the EyeCloudAI-CDK developer kit with three swappable sensor modules (2 MP rolling shutter, 2 MP global shutter, 8 MP), enclosed USB AI cameras, IPC cameras and a temperature-measurement module. One OpenVINO workflow, one SDK — and one of its deployments measured footfall, dwell and queues for a European retail intelligence provider. |
🔒 The camera that never remembers Every video frame is analyzed inside the camera and the edge box and then discarded. What leaves the device is anonymous metadata — counts, dwell times, queue lengths. No images or videos are stored, streamed or exported. |
|
🧩 OpenVINO inside, models swappable in under a minute Intel Movidius Myriad X VPU (up to 4 TOPS) runs Intel's OpenVINO toolkit natively on the camera. Customers import models from the Open Model Zoo or their own training — object detection, person counting, pose, custom nets — and swap the running model through the open-source OpenNCC View tool in under a minute. |
🧰 Your algorithms, our hardware The customer's own analytics stack was integrated into the EyeCloud edge box through the OpenNCC SDK — one API covering USB AI cameras, IPC cameras and temperature-measurement variants. EyeCloud supplies the cameras and the box; the customer owns the intelligence. |
Definition. OpenNCC (Open Neural Compute Camera) is EyeCloud AI's programmable vision-AI platform: an Intel Movidius Myriad X VPU paired with a camera sensor and an edge box, running Intel's OpenVINO toolkit directly on the device. AI models are not baked in — they are loaded, changed and re-loaded by the customer. The platform ships with an open-source configuration tool (OpenNCC View), a unified SDK, open hardware documentation and a developer community.
Where a conventional smart camera runs whatever the vendor embedded at the factory, an OpenNCC device is a blank vision appliance: the customer decides which OpenVINO models run on it, and can change them in the field in under a minute.
| OpenVINO property | What it means in a product |
|---|---|
| Inference runs on local hardware (VPU/CPU/GPU) | Real-time results with no cloud round trip — works behind the store's firewall, even offline |
| Imports models from PyTorch, TensorFlow, ONNX, PaddlePaddle into an optimized Intermediate Representation | Teams keep their existing training stack; deployment is a conversion step, not a rewrite |
| Open Model Zoo: pre-trained, optimized models for detection, classification, pose and more | Standard perception (person detection, counting) starts from proven models instead of bespoke training |
| FP16/INT8 optimization for low-power VPU execution | Full-fps multi-model analytics inside a few watts — the box fits anywhere, needs no active cooling |
Both Open Model Zoo models and fully custom customer networks run on the platform — including two models chained in series, the output of one feeding the next, for complex single-device inference.
| Device | Form factor | Key specs | Status |
|---|---|---|---|
| OpenNCC core board (DK) | 38 × 38 mm, 9 g | Myriad X VPU, up to 4 TOPS, 8 Gbit LPDDR4, USB Type-C 3.0, GPIO/I²C/SPI/UART expansion | EOL |
| EyeCloudAI-CDK developer kit | Core board + swappable sensors | Three M12 sensor modules: 2 MP rolling shutter, 2 MP global shutter (120 fps @ 1 MP), 8 MP (3872 × 2180); H.264/H.265 encode; Raspberry Pi compatible | EOL |
| OpenNCC Knight USB AI camera | Enclosed USB camera | 1080p @ 25 fps, 70° FOV, runs OpenVINO models on-board; multi-camera deployment over USB hub | EOL |
| OpenNCC IPC camera | Industrial housing | Standalone IP-camera form factor with on-device analytics, ONVIF-oriented integration | EOL |
| OpenNCC IR384 temperature camera | Thermal module | Temperature measurement layer in the SDK alongside ISP and video | EOL |
 |
 |
 |
Side profile and interface detail: HDMI, GPIO, power, USB and RJ45 — everything runs on-device.
 |
 |
 |
HDMI and USB on the side, tripod-mountable — the deployment above pairs the IR384 with an Intel NUC host.
The OpenNCC device family has reached end of life — the deployments below ran on it, and the same architecture (edge box + cameras + on-device OpenVINO inference) is available today as a complete EyeCloud AI solution.
What ties these devices together is the platform underneath: the same Myriad X VPU, the same OpenVINO runtime, the same SDK calls — whether the camera is a bare 38 mm board on a developer's bench, a USB camera on a monitor, an IPC unit on a store wall, or a thermal module at an entrance. The form factor is chosen per application; the intelligence stays portable. Deployments of the platform range from fever screening at a California surgery center to the retail case described below.
The customer operates a retail intelligence platform for grocery, fashion and other store formats across Europe: it measures footfall, dwell time, queue buildup and conversion, and turns those signals into staffing and layout recommendations on a live dashboard. Their requirements for the sensing layer were non-negotiable:
| Layer | Deployed |
|---|---|
| Capture | EyeCloud camera modules on the OpenNCC platform (fixed-mount, M12/C-mount lens options), aimed at entrances, aisles and checkout queues |
| Edge AI compute | EyeCloud edge box with Intel Movidius Myriad X VPU running the OpenVINO runtime |
| Intelligence | The customer's own detection and tracking models, integrated through the OpenNCC SDK and loaded via the OpenVINO toolchain |
| On-device output | Anonymous event metadata only — counts, dwell and queue statistics. No frames retained |
| Cloud / dashboard | The customer's existing retail intelligence platform, consuming events — never video |
When a store chain's legal team — or a national data protection authority — asks the question, the answer is architectural: there are no images to subpoena, because none exist. No footage leaves the device; no frames are stored on it; the only data that crosses the network is aggregate counts a person cannot be re-identified from. GDPR's data-minimization and privacy-by-design obligations are not policies layered on top of this product — they are consequences of how it processes data. The provider can state flatly what surveillance-era competitors cannot: no images or videos are stored.
| Outcome | Status |
|---|---|
| Images or videos stored, on-device or in cloud | None — frames discarded after on-device inference |
| Personal data crossing the network | Aggregate anonymous statistics only |
| Store installation effort | Compact plug-and-play sensing device, in service within minutes, no infrastructure changes |
| Algorithm ownership | 100% customer models running on OpenNCC hardware |
| Deployment footprint | Store formats across grocery, fashion and other sectors in Europe |
| Consideration | OpenNCC edge analytics | Cloud-streaming camera |
|---|---|---|
| Privacy / GDPR | Frames never leave the device — nothing stored, nothing to breach; data minimization by architecture | Raw video uploaded and retained — per-deployment justification, retention schedules and DPIA overhead |
| Bandwidth & cost | Kilobytes of event data | Continuous video streams per camera, per store, per day |
| Latency | Real-time local inference — alerts fire while the queue is still forming | Round-trip delay; insights arrive after the moment |
| Store IT impact | No video load on the store network | IT objections at every rollout |
| Algorithm ownership | Customer models run on customer hardware | Vendor-locked analytics pipelines |
A programmable vision-AI platform from EyeCloud AI: Intel Movidius Myriad X VPU hardware (cameras and an edge box) running Intel's OpenVINO toolkit on-device, with an open-source configuration tool, a unified SDK and open documentation. Customers load and swap their own AI models instead of consuming a fixed vendor feature set.
Video is analyzed inside the device and discarded immediately; only anonymous aggregate statistics leave it. There is no footage to store, retain or disclose — data minimization by architecture.
Yes — that is the platform's purpose. Models trained in PyTorch, TensorFlow, ONNX or PaddlePaddle convert to the OpenVINO format and run on the Myriad X VPU. The Open Model Zoo supplies pre-trained starting points, and the OpenNCC View tool swaps the deployed model in the field in under a minute.
A 38 × 38 mm core board, the EyeCloudAI-CDK developer kit with three swappable sensor modules (2 MP rolling/global shutter, 8 MP), enclosed USB AI cameras, IPC-style cameras and a temperature-measurement variant — all reachable through one SDK.
It was. With Intel's Neural Compute Stick 2 discontinued, the OpenNCC platform ran the same OpenVINO workflow in camera and box form factors. The platform is now EOL as well — EyeCloud AI's current edge AI solutions (AI computing boxes + GMSL2/USB cameras) carry the same approach forward.
OpenNCC has reached end of life — the edge vision expertise behind it hasn't. If your product needs an edge AI vision solution, EyeCloud AI provides the complete package: GMSL2 and USB cameras, AI computing boxes, frame grabbers and the accompanying software stack — integrated, tested and shipped as one solution, with ODM customization for your application. Your algorithms run on the edge and your data stays there. Contact us to scope your project.
