⚠️ Important Notice
This model is provided for evaluation and development purposes only. It is not validated for and must not be used in clinical, diagnostic, or production settings. See #use-and-limitations and #License.
Model Information
Description
latch-detect is a single-stage object detection model that detects the latch clips on the corners of a box compartment. It uses the same MobileNetV2 + FPN + anchor-free head architecture as the companion person and patient detection models, retrained to recognize latch hardware. Downstream logic infers the state: one or more detected clips indicates the hood is closed/latched; zero detections indicates it is open/unlatched. Asymmetric hysteresis (2 consecutive open readings to open, 3 to close) is applied in the application layer for stability. This model is an internal development and evaluation tool produced as part of Intel's NICU Warmer reference design and has not undergone clinical validation. In this reference application, the detection of the latch clips on the box is designed to mimic the workload of an AI model operating in a real neonatal scenario where it would be able to recognise the presence of latch clips on a NICU Warmer.
Intended Use
This model is intended for use by software developers and researchers evaluating AI-assisted equipment-state monitoring on Intel hardware. It demonstrates detection of small mechanical features (latch clips) in video as a proxy for hood-closed/open status. latch-detect itself does not provide any medical functionality, nor is it intended to process or interpret medical data for a medical purpose. Developers are responsible for independently validating and adapting latch-detect for their specific use case. It must not be used in live clinical environments or relied upon for patient-safety decisions.
Technical Specifications
| Attribute | Detail |
|---|---|
| Architecture | MobileNetV2 backbone + FPN neck + anchor-free detection head |
| Parameters | ~5–7M (FP32) |
| Input | 992×800 RGB image (W×H), float32, normalized [0,1], NCHW |
| Output | [N, 5] — bounding boxes (x1, y1, x2, y2, confidence), post-NMS, pixel coords at input resolution; application layer counts detections to classify hood state as "open" or "closed |
| Training hardware | Intel Ultra Core |
| Framework | PyTorch (mmdetection) → OpenVINO IR FP32 |
Training Data
The model was trained using images of a box-like compartment taken from above. These images were created by the development team for this purpose. Within the box compartment (representing the NICU Warmer), numerous images show different configurations of scenarios i.e. the presence or absence of a plastic figure (representing the “patient”), the presence or absence of a hand (representing the “caretaker”) and the presence or absence of latch clips (representing the NICU Warmer hood latch). The model is trained to recognise the presence of latch clips within this environment. Once trained, you can test the performance of the AI model and by extension the hardware in real-time to mimic production use. Whilst the training process allows for simulation of production workloads, the fact that the training data has no clinical nexus demonstrates that the NICU Warmer reference application was not intended for use within a production, clinical environment.
Evaluation
Evaluation was limited to confirmation that the model can recognise the presence of latch clips in the environment described above. No formal evaluation of performance was undertaken. It is recognised that reference AI workloads may not replicate real, production workloads
Use and Limitations
Permitted Uses
- Evaluation and benchmarking of Healthcare and Life Sciences AI workflows
- Research and development
- Academic study
Prohibited Uses
- Clinical or diagnostic use
- Production deployment
- Use with live patients or real patient data
Known Limitations
- No evaluation of performance undertaken
- No evaluation of whether reference AI workloads will replicate real, production workloads
License
The use of latch-detect is governed by the Intel Limited Internal Research & Development Use License Agreement. By accessing this model, you agree to the license terms.