Rubik Pi3 how to get inference time object detection

Issues
1

Hi, I want to do a benchmark of the inferencing time of various YOLO models on the Rubik Pi3. I’m having trouble getting to the inference times. I send detections via appsink to a Python script that calculates the inference time based on the time difference between the sent detection buffers. The problem is that by not setting a frame-rate in the pipeline I always get an inference time of 33ms, regardless of the size of the model. On the other hand, when I increase the frame-rate, the inference time increases by a few ms. Is there any way to effectively get to the inference time to calculate FPS from the buffer information?
I am uploading my pipeline below:

“”"
qtimetamux name=metamux ! queue ! qtioverlay ! queue ! waylandsink sync=true fullscreen=false
qtiqmmfsrc name=qmmf camera=0 !
video/x-raw(memory:GBM),format=NV12 !
tee name=split ! queue ! metamux.
split. ! queue name=detect_q ! qtimlvconverter ! queue !
qtimlsnpe delegate=dsp model={args.model_path} layers=“</model.22/Mul_2, /model.22 /Sigmoid>” !
queue ! qtimlvdetection name=detection threshold={args.detection_threshold * 100} results={args.max_detections} module=yolov8 labels={args.labels_path} !
text/x-raw, format=utf8 !
queue ! tee name=split2
split2. ! queue ! metamux.
split2. ! queue ! appsink name=sink emit-signals=true sync=true drop=false
“”"

Greetings

2

Dear customer,we are currently syncing this issue internally.

3

Dear customer
Before and after the function ①, obtaining the timestamps and calculating the difference will provide the exact inference time.

图片
图片1330×949 116 KB

4

Okay I will try it, thank you

5

This is the time of the model inference itself, but what if I want to get the time of total processing of one frame with pre and post processing included

6

Dear customer
What does “total processing time” specifically refer to? Is it the time from when a frame request command is issued until the display is completed (including inference time)?

7

I mean the time from issuing the frame request command to getting the coordinates of the bounding boxes, because let’s say I don’t need a display, which in itself takes some time.

8

Dear customer
One terminal executes object detection, while another terminal executes the following commands to obtain the time of issuing the frame request command:

journalctl -o short-iso-precise -f  | grep process_capture_request

for example:

图片
图片1658×287 139 KB

After function ①, obtain the time of getting the coordinates of the bounding boxes.
图片
图片690×492 81.7 KB