inference time

[zhucheng725]

Hi. I follow this repo

$ roscore

$ rviz -d ./config/show.rviz

$ python livox_detection_simu.py

$ rosbag play *.bag -r 0.1

and my command window shows like this:

det_time(ms) 47.05095291137695
det_numbers 10
det_time(ms) 43.282270431518555
det_numbers 11
det_time(ms) 45.348167419433594
det_numbers 10
det_time(ms) 42.76132583618164
det_numbers 12
det_time(ms) 45.06182670593262
det_numbers 11
det_time(ms) 44.640302658081055
det_numbers 10
det_time(ms) 43.58649253845215
det_numbers 10
det_time(ms) 43.75743865966797
det_numbers 11

Dose it mean that my PC(Nvidia 1060Ti) can run this about 40 ms / per frame when I buy a livox? I also watch my GPU and nothing to be used when I run this demo. Dose it mean this demo do not use GPU? I have installed TF1.14_gpu. Many thanks!

[rosexplorer]

Hello, I tried it on a GTX 1050 and it had an interference time of 50ms. When I tried it with the lidar, the interference time didn't change, but the objects were detected after one minute. I don't know what is the problem, so I bought the Nvidia jetson Xavier NX and the tool isn't working like its supposed to on this computer.

[zhucheng725]

Could you post your terminal warnings or errors here so that we can analyze. You can make some nodes to see how fast it is

[rosexplorer]

I opened a topic with all information on Stack overflow: https://stackoverflow.com/questions/66823106/jetson-xavier-nx-lidar-object-detection-tensorflow-starting-problems

How can I make a node to see how fast it is? I am completely new to ros and deeplearning.

[zhucheng725]

I have seen your log in stack and it shows your NX is out of memory. NX is not better than gtx1050. You can check some info about NX.

[rosexplorer]

But technically the nx has about 6gb for the gpu and the 1050 only has 2gb. The AI performance is also higher (6Tflops on NX and 1,7 TFlops on 1050). Or am I wrong?

[zhucheng725]

If you want , you can have a try to calculate dl demos such as Vgg16 both GTX and NX. NX:(double:1733, single:54, half:27 . gtflops and 1050(double:700 . gflopsg. 1050 has about 700 cuda core and NX has 384.

[rosexplorer]

anyways, the problem will be the same: objects get detected too late. I think it is because the data comes in and gets cached and used one after another. Is there a possibility that the program only takes the data from the lidar which is currently recheived from the lidar?

[Livox-SDK]

Hi zhucheng725,

I think the lag is caused by the pre-process part taking too long, re-write the pre-process part by C++ the lag will get improved.

[Double-zh]

你好，请问如何可视化检测结果

[Double-zh]

我在使用rivz中尝试pointclound2进行可视化，遇到这个问题For frame [livox_frame]: Fixed Frame [map] does not exist