Real time object detection

Autonomous Machines Jetson & Embedded Systems Jetson TX2
1

Hello everyone,
I am new at using the jetson TX2 and I really wanted to try out the object detection HOWEVER I am very lost on how to do this. Ive read several papers but i am still VERY lost. does anybody have a step by step tutorial (also listing the programs that i need to download) on how to achieve object detection real time? I was kind of gravitating towards using Single Shot Multibox method but let me know what you guys think its best. Also if anyone knows of any good/useful websites to do this link them below.

2
3

Hi,

If you want to use SSD, you should build their branch by your own.

But if you can use another algorithm, check these pages:
https://developer.nvidia.com/embedded/twodaystoademo

It will guide you to use detectNet in details.

Thanks.

4

I’ve written a couple of blog posts sharing my experiences running YOLOv2 and Single Shot Multibox Detector (SSD) on Jetson TX2.

[url]https://jkjung-avt.github.io/yolov2/[/url]
[url]https://jkjung-avt.github.io/ssd/[/url]

I also shared my experiences with NVIDIA DetectNet in earlier posts. Feel free to check them out.

5

I shared a python script which could do real-time SSD object detection with either JTX2 on-board camera, USB webcam or IP CAM. Please refer to the following blog post.

[url]https://jkjung-avt.github.io/camera-ssd-threaded/[/url]

6

Hi @jkjung13 .

Thank you so much for your posts, they are very helpful.

I have some questions about object detection with Jetson TX2 that I need to be clarified.

I believe that you use caffe and tensorflow in your code. Have you tried pytorch ? It can get the same speed as tensorflow or caffe ?

Does coding in python affect the performance ? is C++ better ?

I am quite new in this problem so I am very appreciated if someone can help me on this.

Thank you so much in advance.

7

@dienhoa.t,

A while ago I wrote a series of blog posts about how to train a tensorflow object detector with custom data and how to optimize the model with TF-TRT (TensorRT), and then to deploy it onto Jetson TX2. Please refer to the following forum thread for more details.

[url]https://devtalk.nvidia.com/default/topic/1042106/jetson-tx2/how-to-train-a-custom-object-detector-and-deploy-it-onto-jtx2-with-tf-trt-tensorrt-optimized-/[/url]

As you have pointed out, I have tested caffe and tensorflow. But I don’t have experience using pytorch on JTX2 yet, so I’m not able to comment on that.

Regarding python vs C++, I’d say if your application is mainly feeding images to DNN for inference (utilizing the GPU) and without much other pre- or post-processing, then the difference in performance should be negligible. In my opinion, programming in python is preferred because of development speed and ease of maintenance.

8

@jkjung13,

Thank you for your response. I will read more in detail your blog posts.

I don’t have experience on TensorFlow so it would be harder for me than with Pytorch but I will try.

Regards,