Object detection doesn't run on GPU

[fiorano10]

I'm trying to run object detection sample on the GPU but it defaults to using the CPU, why?
'[ INFO ] --------parameters DUMP---------------------
[ INFO ] Pipeline: object
[ INFO ] Inputs: RealSenseCamera,
[ INFO ] Outputs: ImageWindow, RosTopic, RViz,
[ INFO ] Inferences:
[ INFO ] Name: ObjectDetection
[ INFO ] Model: /opt/intel/computer_vision_sdk/deployment_tools/intel_models/person-vehicle-bike-detection-crossroad-0078/FP32/person-vehicle-bike-detection-crossroad-0078.xml
[ INFO ] Engine: GPU
[ INFO ] Label: to/be/set/xxx.labels
[ INFO ] Connections:
[ INFO ] ObjectDetection->ImageWindow
[ INFO ] ObjectDetection->RosTopic
[ INFO ] RealSenseCamera->ObjectDetection
[ INFO ] Common:
[ INFO ] camera_topic:
[ INFO ] custom_cpu_library:
[ INFO ] custom_cldnn_library:
[ INFO ] enable_performance_count: 0
getting instance
return instance
[ INFO ] device_FACE:CPU
[ INFO ] model_FACE:/opt/intel/computer_vision_sdk/deployment_tools/intel_models/person-vehicle-bike-detection-crossroad-0078/FP32/person-vehicle-bike-detection-crossroad-0078.xml
[ INFO ] device_AG:CPU
[ INFO ] model_AG:
[ INFO ] model_HeadPose:
[ INFO ] device_HeadPose:CPU

[chaoli2]

@fiorano10
The default yaml for people detection combines 4 inference instances for face, emotion, age & gender and head pose respectively. Each inference will involve one inference engine like CPU or GPU.

If you'd like to do object detection, just keep one inference instance and remove others. Here is one object detection yaml example for your reference.

[fiorano10]

I already have the object detection running on CPU, but even if I switch it to GPU in the yaml, the inference still defaults to CPU as I mentioned above. When I run the object detection the CPU usage jumps to 600%.

[chaoli2]

@fiorano10
Will you show me your yaml file?
According to your provided log, the inference engine launched 4 inferences at the same time. One is running object detection on GPU, and another three inferences are running detections on CPU, such as face, age&gender and headpose detection.

[fiorano10]

This is what yaml file looks like:
`Pipelines:

• name: object
  inputs: [RealSenseCamera]
  infers:
  • name: ObjectDetection
    model: /opt/intel/computer_vision_sdk/deployment_tools/intel_models/person-vehicle-bike-detection-crossroad-0078/FP32/person-vehicle-bike-detection-crossroad-0078.xml
    engine: GPU
    label: to/be/set/xxx.labels
    batch: 16
    outputs: [ImageWindow, RosTopic, RViz]
    confidence_threshold: 0.2
    connects:
  • left: RealSenseCamera
    right: [ObjectDetection]
  • left: ObjectDetection
    right: [ImageWindow]
  • left: ObjectDetection
    right: [RosTopic]

OpenvinoCommon:`

[chaoli2]

@fiorano10
The yaml is OK.
But I have not met this issue from my side. The runtime inference engine is exactly what I specified in the yaml...
The provided log is the output of above yaml?

[fiorano10]

Yup, that's the output of the above yaml file. when I run face or emotion detection, it runs fine on the GPU but not object detection

[fiorano10]

You need to add FLAGS_d = pipelines[0].infers[0].engine; in /ros_openvino_toolkit/sample/src/sample_pipeline_object.cpp

[songshan0321]

I have written a easy-to-use ROS wrapper for OpenVINO R3 Inference Engine, with a step-by-step instruction! =) https://github.com/songshan0321/ros_vino