Skip to content
SENSORAIA / DOCUMENTATION

Choose And Run An AI Model

Sensoraia can run compatible ONNX models on the CPU, using the same native inference support across its sensor applications. You choose a separate model for each job: a vehicle detector, sound classifier and sampling planner do not interpret the same outputs.

AI terms explained | Spottercam | Diplacam

Choose A Model

Start with a model exported for the application you want to use. Keep its graph, descriptor, labels and any deployment metadata together. The filename extension alone does not establish compatibility.

Application The Model's Job
Spottercam Return image boxes, classes and scores
LOOMcam Provide features or an explicitly declared return vector; a trained homing model is not included
EARcam Classify the declared microphone sample window
LWIRcam Classify declared display pixels or supported temperature samples
Diplacam Predict one denoising step used by the native measurement planner

Each application reads its own settings file, such as modules/spotter_cam/spottercam_settings.ini. A typical ONNX assignment is:

inference_backend=onnx
model=models/assignment/model.onnx
model_descriptor=models/assignment/model.ini
onnx_threads=1
inference_timeout_ms=2000

Paths in these settings are relative to the INI file, not your terminal. Use --settings /absolute/path/settings.ini to select another assignment. Keep the supplied descriptor with the graph: it records the expected input preparation, tensor layout and graph identity.

Spottercam and LOOMcam can also use a compatible IMX500 .rpk with the physical AI Camera. inference_backend=auto selects ONNX for .onnx or IMX500 for .rpk on hardware; during PC simulation or saved-image CPU analysis, Spottercam automatically probes sibling vehicle_yolo11n_int8.onnx and vehicle_yolo11n_fp32.onnx exports in models/ when model points to models/nn_vehicles.rpk. CPU inference does not require the camera's neural processor.

Check The Assignment Before Capture

From the Sensoraia root, for Spottercam:

make -j32 spottercam CAMERA_CAPTURE=0
bin/spottercam --check-model --settings /absolute/path/spottercam_settings.ini

CAMERA_CAPTURE=0 builds without physical camera support. It still runs real model inference. Omit that switch when you need live camera capture. The model check does not take a photograph or establish detection accuracy.

Try A Saved Picture

Use an original image representative of the model's intended camera input:

bin/spottercam --analyze /absolute/path/photo.jpg \
  --settings /absolute/path/spottercam_settings.ini --sequence 1

Inspect the result beside the original and read any reported rejection. Ordinary analysis leaves the input unchanged and writes the application's result sidecar. The module guide explains its result fields and limitations.

Saved images, MIMIC views and supported live cameras can feed the same model processing. Their provenance remains different: a real neural result from a simulated image is still not a real-flight measurement.

Understand Empty Results And Errors

Result What To Do
Model accepted, no objects found Keep the empty result; do not invent targets
Descriptor/hash mismatch Restore the graph and descriptor from the same export
Unsupported tensor layout or preprocessing Obtain a compatible export; do not change labels merely to silence the check
Runtime library unavailable Use the matching packaged runtime for this processor and release
Inference timeout or failure Read the diagnostic and stop relying on fresh model results until resolved

Explicit PC demos can use labelled random observations under their documented fallback rules. That is a simulation choice, not normal production error handling. A present invalid model must not silently become a random detector.

Put The Assignment On MORA

Build and package on the PC using the deployment workflow. The release needs matching runtime libraries, selected model files and their settings. Review the module's package selection before activation; a successful local image run does not measure MORA's shared memory use or sustained processing rate.

Evaluate the model on independent examples resembling its intended use. Scores are not automatically calibrated probabilities, and changing the camera, height, lighting or sound environment can change performance.