yehengchen / Object-Detection-and-Tracking

Object Detection and Multi-Object Tracking
MIT License
1.85k stars 753 forks source link

Tiny-yolov4 + DeepSort #79

Open MuhammadAsadJaved opened 4 years ago

MuhammadAsadJaved commented 4 years ago

Hi, 1-Is there any way to use Tiny-yolov4 + deepSort?
2- in the NVIDIA NX yolov3+DeepSort is very slow. Is it possible to use TensorRT to optimize yolov3 or v4 and then combine with DeepSort to make it faster? 3-Any other suggestion to run it faster in NX?

ewuxjun commented 4 years ago

Hi,i use the tensorrt to test yolov3(tensorrt)+deepsort , the performance is 5fps, what about you ?

MuhammadAsadJaved commented 4 years ago

@ewuxjun Hi buddy, I am trying to run it in NX, It still have some problems. Which project are you using for Jetson and which device are you using?

ewuxjun commented 4 years ago

maybe it is the same as your project, the device is also jetson Xavier nx

MuhammadAsadJaved commented 4 years ago

OH. Great, let me try then I will share the results with you. BTW I am using this project,

https://github.com/Stephenfang51/tracklite

I am confused about one thing, the only YOLOV3 TRT can run in real-time in NX. Some people said in the YOLO+DeepSort most of the time is consumed by YOLO so that's why YOLO+DeepSort is slow. But if we converted YOLO to TRT and then combined with DeepSort then why it is still very slow? Do you have any idea?

ewuxjun commented 4 years ago

yeah, I have encountered the same problem, yolov3 can run in real-time in Xavier nx ,as you say, yolov3+deepsort is so slow, I don't know these two modules influence each other, so now I am searching for resnet+deepsort, we just know yolov3 need much performance

MuhammadAsadJaved commented 4 years ago

@ewuxjun let's try it. Is it possible to connect over WeChat as we are working on the same problem, it might be helpful?

If Yes then my WeChat is: 13621297680

ewuxjun commented 4 years ago

ok, I have added friends, you can check it

MuhammadAsadJaved commented 4 years ago

Received. Thanks