OpenVINO Detector
OpenVINO 适合 Intel 平台,可在 CPU、集成 GPU 以及支持的 Intel 加速硬件上运行模型。它既可用于传统 OpenVINO IR 模型,也可作为部分 ONNX 工作流的后端。
SSD 示例
detectors:
ov:
type: openvino
device: CPU
model:
width: 300
height: 300
input_tensor: nhwc
input_pixel_format: bgr
path: /openvino-model/ssdlite_mobilenet_v2.xml
labelmap_path: /openvino-model/coco_91cl_bkgr.txt
device 怎么选
CPU:兼容性最好,适合验证模型。GPU:使用 Intel GPU 推理,需容器可访问/dev/dri。AUTO:让 OpenVINO 选择可用设备,适合一般部署,但故障排查时建议先指定明确设备。
视频解码与 OpenVINO
同一块 Intel iGPU 可以同时承担 FFmpeg VAAPI/QSV 解码和 OpenVINO 推理,但资源是共享的。摄像头较多或模型较大时,要同时观察解码和推理延迟,而不是只看某一个百分比。
模型不匹配的典型表现
- 启动时报 shape/layout 错误:检查 300×300、NHWC/NCHW。
- 框位置正确但标签错误:检查 labelmap 类别顺序。
- 检测结果很差:确认 BGR/RGB、dtype 和预处理假设。
调试建议
先使用 CPU 把模型正确性验证通过,再切 GPU/AUTO;这样可以把“模型配置问题”和“硬件访问问题”分开定位。
OpenVINO detector
OpenVINO is a strong fit for Intel systems and can use CPU, integrated GPU, and supported Intel accelerators. It is commonly used with OpenVINO IR models and can also participate in ONNX workflows.
SSD example
detectors:
ov:
type: openvino
device: CPU
model:
width: 300
height: 300
input_tensor: nhwc
input_pixel_format: bgr
path: /openvino-model/ssdlite_mobilenet_v2.xml
labelmap_path: /openvino-model/coco_91cl_bkgr.txtChoosing a device
CPU: easiest baseline for model validation.GPU: Intel GPU inference; the container needs access to/dev/dri.AUTO: lets OpenVINO select hardware; convenient in production, less explicit while troubleshooting.
Decode vs inference
The same Intel iGPU can handle FFmpeg VAAPI/QSV decode and OpenVINO inference, but both compete for shared resources. Watch decode stability and inference latency together.
Typical mismatches
- Shape/layout errors → width/height or NHWC/NCHW.
- Correct boxes but wrong labels → label map order.
- Poor detections → RGB/BGR, dtype, or preprocessing mismatch.
Debug method
Validate the model on CPU first, then move to GPU/AUTO. This separates model metadata errors from device-access problems.
Detector OpenVINO
OpenVINO es ideal para plataformas Intel y puede usar CPU, iGPU y aceleradores compatibles.
Ejemplo SSD
detectors:
ov:
type: openvino
device: CPU
model:
width: 300
height: 300
input_tensor: nhwc
input_pixel_format: bgr
path: /openvino-model/ssdlite_mobilenet_v2.xml
labelmap_path: /openvino-model/coco_91cl_bkgr.txtDispositivo
CPU: base sencilla para validar.GPU: requiere acceso a/dev/dri.AUTO: OpenVINO selecciona el hardware.
Decode e inferencia
La misma iGPU puede decodificar vídeo y ejecutar inferencia, compartiendo recursos.
Errores típicos
- Shape/layout → dimensiones o NHWC/NCHW.
- Etiquetas erróneas → labelmap.
- Detección pobre → RGB/BGR, dtype o preprocesado.
Depuración
Valide primero en CPU y después cambie a GPU/AUTO para separar errores de modelo de problemas de acceso al hardware.
