search

dusty-nv / jetson-inference

Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.
https://developer.nvidia.com/embedded/twodaystoademo
MIT License
7.87k stars 2.98k forks source link

how to increase the FPS on Peoplenet model? #1739

Open saisatheesh5 opened 1 year ago

saisatheesh5 commented 1 year ago

Hi All,

I have Advantech jetson nano (MIC-710AI). I am using the PEOPLENET Model and I have connected 2 cameras to that for live detection. Now I have received only 3FPS. For my requirement, I need at least 25 to 30 FPS. How can I increase my FPS. Please provide you suggestions to improve that. And Previous forum observation I have tried the below comments.

1) sudo nvpmodel -m 0 2) sudo jetson_clocks

Above two commands I ran but after that also I don't have any improvement on that.

Thanks in Advance!

dusty-nv commented 1 year ago

@saisatheesh5 what program are you using? Which camera? Are you able to stream one camera alone with video-viewer at > 3FPS?

saisatheesh5 commented 1 year ago

Hi dusty,

Thanks for the response and valuable time.

I am using detectnet program with peoplenet model. I am using Upcamera 2MP. it is made by intel.

Yeah, I tried with one camera alone with the same program still i received the same.

when I tried with SSD-mobilenet-V2 model, on that time i have received 14 to 16 FPS.

up camera datasheet.pdf

Thanks!

dusty-nv commented 1 year ago

@saisatheesh5 my guess is that since Jetson Nano is unable to do INT8, and PeopleNet is a higher-resolution model that becomes fast using INT8 quantization, it is slower on FP16-only Nano. You can try the other support TAO detectNet_v2 models and see if the performance scales.

saisatheesh5 commented 1 year ago

Hi dusty-nv,

Thanks for your response.

If I have installed the deepstream is any possible is there to increase the performance?

Please suggest any idea to detect only the human in a FOV using jetson nano.

My main purpose was a need to inspect a specific area if any humans are there or not. So please suggest which model I can use for that with Jetson Nano. I have tried SSD-mobile net-V2 with person class but most of the time it detects the robot as a human from various angles.

Thanks

dusty-nv commented 1 year ago

DeepStream typically provides the most optimized performance yes, and works with the TAO models. Here are Python examples for DeepStream: https://github.com/NVIDIA-AI-IOT/deepstream_python_apps

saisatheesh5 commented 1 year ago

Hi @dusty-nv,

Thanks for your suggestion. I have found the more accurate Yolo model. I tested with my trail videos it's working perfectly so please help me to implement that on the jetson nano from Advantech(Mic-710ai).

I have mentioned the link below which I followed to use the Yolo human detection model.

Thanks!