About the frame rate problem of crossroad_camera_demo

[STATEABC]

I am using FPGA AI Suite to run crossroad_camera_demo in OpenVINO.

The models I use are:
person-vehicle-bike-detection-crossroad-1016
person-attributes-recognition-crossroad-0230
person-reidentification-retail-0288

If all three models are run on the CPU, the average frame rate is around 45
If the main recognition model 1016 is run through HETERP:FPGA,CPU, and the other two are run through the CPU, the frame rate will not exceed 20

This is very confusing and distressing to me, why is the frame rate reduced?

And do attributes-recognition-0230 and reidentification-retail-0228 not support HETERO:FPGA,CPU? No matter which arch and sof I switch, there will be an error

Any help would be greatly appreciated, thank you!

[Peh_Intel]

Hi STATEABC,

For your information, the three models that you are using are not listed in the supported models of FPGA AI Suite.

Models that having activation operations that are not supported on the primary device are split by Hetero into multiple sets of subgraphs which leads to suboptimal execution.

This may cause the data sending back and forth many times during one inference, which resulting in heterogeneous mode takes more time than under normal execution.

Regards,

Peh

[Peh_Intel]

Hi STATEABC,

This thread will no longer be monitored since we have provided answer. If you need any additional information from Intel, please submit a new question. 

Regards,

Peh