Model Compatibility and Output Formats

This page summarizes what runs on the reCamera Pro NPU (Rockchip RV1126B, 3 TOPS, mixed INT8/INT16) and what the inference pipeline outputs. Use it as a quick check before you invest time converting or training a model.

Supported model format

Item	Value
On-device model format	RKNN (.rknn), compiled for RV1126B
Required target platform	target_platform='rv1126b' when building with RKNN-Toolkit2
Conversion toolkit	RKNN-Toolkit2 2.3.2, host-side (x86_64 Linux or WSL 2, Python 3.6–3.12)
Runtime alignment	Keep the Toolkit version aligned with the RKNN Runtime shipped in the reCamera Pro firmware
Precision	FP16 (non-quantized baseline) and INT8 (quantized with a calibration dataset)
Source formats	ONNX (via RKNN-Toolkit2 or SenseCraft conversion); Ultralytics YOLO checkpoints (.pt) can export directly to RKNN

• Do not use rknn-toolkit-lite2 for conversion — it is a device-side runtime library, not a converter.
• A model built for a different Rockchip SoC (e.g., RK3566/RK3588) will not run on RV1126B.

Conversion walkthroughs: RKNN-Toolkit2 Conversion (host, scriptable) and SenseCraft ONNX-to-RKNN (browser-based, no code).

Built-in and tested model families

The firmware ships with built-in detection models, and the Web UI Model Management list exposes these attributes per model:

Field	Typical values
Framework	RKNN
Algorithm	YOLO, nanodet
Type	Detection

Tested/bundled use cases include person detection, hard-hat detection, construction-safety detection and vehicle detection.

Input contract requirements

Your ONNX model must have a known, static input contract before conversion. RKNN applies (input − mean) / std once — if normalization is already inside the ONNX graph, use identity values (mean=[0,0,0], std=[1,1,1]).

Item	Requirement
Input shape	Static, batch-1 preferred (e.g., [1, 3, 640, 640])
Layout	NCHW or NHWC, matching the exported graph
Channel order	RGB vs BGR must match training — a swap severely reduces accuracy
Resize policy	Stretch, crop or letterbox must be identical at runtime
Unsupported operators	Re-export the ONNX model or use a semantically equivalent supported op

Inspect the contract with Netron or model-inspect before converting. Details and troubleshooting: RKNN-Toolkit2 Conversion.

Custom classes and post-processing

Uploaded models are configured in the Web UI (Upload and Configure an RKNN Model, Configure Detection):

• Detection Categories — set output class names manually or bulk-import a category list from a TXT file; inference results then display your own labels
• Post-processing parameters — IOU (NMS threshold), Confidence (detection threshold), max_obj (max objects per frame)

Inference output formats

Real-time inference output (Web UI monitoring, logs) uses these fields:

Field	Description
timestamp	Inference result timestamp
task_type	Current task type
class_id	Detected class ID
class_name	Detected class name
score	Detection confidence
bbox	Bounding box coordinates
detection_count	Number of objects detected in the current frame

Results can be pushed to external systems over three channels, with configurable output template and task type:

Channel	Guide
HTTP	Send Detection Results over HTTP/UART
UART	Send Detection Results over HTTP/UART
MQTT	Send Detection Results over MQTT

Recordings can also be triggered directly by inference results (categories, confidence range, trigger areas): Configure Event Recording.

Native (C/C++) inference

For custom applications outside the Web UI, models are loaded through the RKNN Runtime C API. Cross-compilation setup and a minimal application: SDK Setup, Native Development. An AI-agent-assisted workflow with a C++ template: Develop with AI Coding Agents.

Sound models

Sound event detection uses a separate pipeline: models are visualized, trained and switched in the Sound Lab (Web UI), and selected sound categories can trigger recording. See Train a Sound Model and Sound-Triggered Capture.

Tech Support & Product Discussion

Thank you for choosing our products! We are here to provide you with different support to ensure that your experience with our products is as smooth as possible. We offer several communication channels to cater to different preferences and needs.