Real-Time Object Detection in 10 Lines of Python Code on Jetson Nano

[@talhakhan3010]

Actual tutorial start at 12:25

[@TrevorLoftus]

bless you

[@pathikghugare]

Thanks 😂😂

[@dellavita3463]

A real Life saver

[@SomaliCoastguard]

Very informative video that has me thinking about using a Nano as the centre of an intelligent alarm system.

[@kwang-jebaeg2460]

Oh my god .. Please give us your excellent explanations more and more

[@kevin_delaney]

A bit!? 😆 my nano is running on an SSD thanks to JetsonHacks and it's actively cooled....you sped that up more than "a bit" 😂 It also took just as long the first time I ran detectnet-console. That worked flawlessly. You should make a video for Yolov3 as well (with cuda and/or opencv), because with that you have like everything covered haha. I'm so hooked!:) thank you!

[@SamAndHenryPlaz]

I used an old Canakit usb wifi and a 16gig microsd (should probably have used a 32gig) card when starting with the Nano last week. I think I waited 2 hours for the default models to download. Plus I didnt have enough space to download PyTorch Then this weekend I added a hardwire ethernet and a usb hard drive and did "pivot the root" from jetsonhacks. Now my Nano life is so much better since it only took < 10min to download all 2.2gig of the models.

[@도전100-h8d]

Very good information and lecture. Thanks for your effort for this video and your job!

[@AWildNoodle]

Really great teacher! I cant wait to get my hands on mine!

[@tacpreppers4906]

just pulled the trigger on a nano, brilliant!

[@marcinkovalevskij5820]

10 lines of python code that you can see, + 10k of c++ code that you can't see

[@maksimmuruev423]

Of course, python has nothing to do with this. Python is slow as hell, cant process video itself especially on small devices. Such example can be with JS or whatever language has wrapper.

[@rodrigoruiz4345]

Thanks for sharing!! I was wondering if any type of camera could be recommended for this application.

[@drtristanbehrens]

This is an EXCELLENT tutorial! Thanks for sharing!

[@sohailawan77]

Please upload tutorial about how to count objects or vehicles in a real time using NVIDIA jetson nano.

[@Jeonghunlee911]

Thanks for sharing information with us, It's easy to follow this tutorial.

[@איילהורוביץ]

Super cool!! Is there some easy way to use your project and fine tune (probably by transfer learning) the detection for specific classes?

[@randommm-light]

Very nice demo. Look forwards to more. I’m curious to make a very minimal bench to play with all this.. what is the essential budget Jetson starter kit? What do I need alongside to make it all happen easily.. starting from scratch and building up.. thx for good work!

[@kevindelnoye9641]

Monitor, keyboard, mouse , Ethernet cable, micro SD with enough memory I recommend 64gb but 32 should be comfortable aswell. You can skip the monitor if you decide to work headless although sometimes a setup with monitor is really useful for displaying results

[@randommm-light]

Kevin Delnoye I get ya on what the jetson needs hooked to it.. I’m all for good monitor, etc.. but what do I need on bench top to get THAT configured and rolling.. Assume starting from iPad.. I need a small ecosystem or can I work w Nano itself and roll uphill..

[@kevindelnoye9641]

@@randommm-light an pc/laptop to flash the SD card with jetpack. After that I don't think you need anything else, optionally a USB camera for testing models

[@drone_video9849]

22:55 it would have been scary / awesome if it detected itself - could have been self-aware...

[@israelip]

If you train it for that, it will.

[@VndNvwYvvSvv]

I don't think you have any idea how computers work, let alone GANs or CNNs.

[@Deepsim]

Thank you for this great tutorial! But I was wondering how to use my own trained model, for example named "my-ssd-mobilenet"?

[@braydenmoore3101]

Awesome tutorial thanks man

[@rockatang2189]

Hi. I followed the steps until make. When I used "make" command, it says "No targets specified and no makefile found." How could I fix this?

[@TheRecep27]

Hello . Can the Yolov7 model run after making these settings? Thank you very much.

[@shubhamcyborg9218]

Make tutorial to train custom datasets and also tell how to remove other datasets to increase fps !!

[@dusty-nv]

Here is a tutorial for training custom image classifiers: github.com/dusty-nv/jetson-inference/blob/master/docs/pytorch-transfer-learning.md For tutorials on re-training custom object detection model, see this post: devtalk.nvidia.com/default/topic/1070225/jetson-nano/digits-or-somthing-else/post/5421938/#5421938

[@bvs7415]

Which jetson nano is best for OpenCv projects??? Jetson nano Orin or any other please reply....

[@ApolloVerified]

I would recommend using Orin as the former jetson nano doesn't get software updates anymore

[@bao_luong]

Please make tutorial about detect custom object on jetson

[@abdiasponce]

x2 !!!

[@rahulmoorkoth2395]

Very informative video. Thanks

[@vvisin13]

If I am in a virtual environment using python 3.11, is there a way to install jetson? I am working with a library that does not work in python 3.8

[@satyamrout1400]

Sir can we use this model to recognize the faces accurately of atleast 60 people at a time? If yes do make a vid on it.....i be very much helpful n yes again an awsm vid....

[@VndNvwYvvSvv]

No, that's insane.

[@fpgamachine]

Excellent it works very well thanks!

[@mychevysparkevdidntcatchfi1489]

Would you change the screen resolution to 640x480 or even to 320x240? I can't read any of the text when your screen is so large.

[@mk677hd]

Cool. gonna play with it soon.

[@owltoe0164]

I followed this tutorial and it works fine, but I only get 5 fps (8 fps with the raspberry pi camera). How do I get to 22 fps like in the video?

[@KevinBocky1]

Thanks, very cool

[@dracleirbag5838]

Using your example where are the files saved? I should delete old files right?

[@el_moustaschbil]

Hello, it seems like some files are missing... When I compile through a make command it appears like: ~/jetson-inference/c/tensorNet.cpp:32:10: fatal error: NvOnnxParser.h: no file or directory of this type #include "NvOnnxParser.h" ... When I take a look on tensorNet.cpp (the file who's in need of the file missing), I can see #include "NvCaffeParser.h" ... #include "NvOnnxParser.h" #include "NvUffParser.h" #include "NvInferPlugin.h" I can't find them and apparently I can't download them because the files aren't on your github. Do you have any idea how I can progress ?

[@dusty-nv]

It seems like you don't have TensorRT installed on your Jetson (TensorRT is included with the JetPack image). Those headers should be found under your /usr/include/aarch64-linux-gnu/ directory If you don't have TensorRT, did you do a custom flash with SDK Manager? If you used SDK Manager, you should choose to install TensorRT on the target. Otherwise, you may want to setup the JetPack image again.

[@adrianoalecrim]

How to run camera ip?

[@chuko2548]

can i optimize with tensorrt a pre-trained tensorflow model ? how can i do that? thank you!!

[@orangehatmusic225]

Sure, just retrain it.

[@sohailawan77]

Please make tutorial on counting of vehicles on a road in real time.

[@dusty-nv]

Hi, I would check the NVIDIA DeepStream SDK which is built for realtime analytics like you mentioned.

[@bschlueter]

Would be cool if the SD card slot also supports ufs cards

[@CodX710]

since its literly booting from it then not realy unless your fine with like 20 minute (idk exactly) boot time :D

[@CodX710]

just noticing that you commented 3 YEARS AGO...............

[@TPA22222]

Really good tutorial 😊. How would I print out the location of the bounding boxes?

[@dusty-nv]

Add these lines of code after net.Detect: print("detected {:d} objects in image".format(len(detections))) for detection in detections: print(detection) You can also access the bounding box and other detection attributes such as ClassID, Confidence, ect by using the members shown here (look under jetson.inference.detectNet.Detection documentation): rawgit.com/dusty-nv/jetson-inference/python/docs/html/python/jetson.inference.html#detectNet

[@TPA22222]

@@dusty-nv thank you 😊

[@akshatjaimini8254]

Is there any way I can combine these with the jetbot repo?

[@afifothman335]

Hello.. Thanks for your video tutorial... I'm already follow this tutorial..but when detect the car, only can detect the from the side of car.. Can't detect from top of car.. How to solve this problem??

[@sajjadakhter7261]

Really good i like it, i was looking for something to identify a person (not just object type), Is thats possible with this library or any idea how it can be done. The reason i am saying person identification not just face recognition is that sometimes its not that easy to get face clear enough to identify a person. Also body has more attributes to compare than just face and i was hoping it may improve accuracy.

[@dusty-nv]

Hi, person recognition or face recognition is typically a 2-stage process where first the coordinates of the person/face are detected (any person, like shown in this tutorial). Then the features from the encoder portion of the network are saved to form a 'profile' of individuals, which are matched against your running database as time goes on. I don't have that later portion in this library, but you could check other implementations online.

[@sajjadakhter7261]

@@dusty-nv Thanks for your response, i am understanding that with this library i can detect where the persons are in, then take that picture and run through some person identification? Any recommended library/product. thanks

[@ZeyadAhmed-om3rs]

how can i take the object detection results to use it in and Arduino ide functions?

[@dheerajdhawan4176]

Has anyone used Jetson nano for custom trained object detection model instead of pretrained models, How was the performance? Is it detecting as fast as shown in demo video? Your feedback would be highly appreciated.

[@AndyKong51]

Any update of doing object detection training on board? Thx

[@AndyKong51]

My application is to track a custom object and output its coordinates. I have done the custom classification based on your tutors. Thx. How can I do the custom object detection? Can I do the training in nano? Slow is fine. If not, do I need to use DIGITS? How to import the model to nano? Thx

[@dusty-nv]

Hi, see this post for some resources on re-training SSD-Mobilenet on custom dataset: devtalk.nvidia.com/default/topic/1070225/jetson-nano/digits-or-somthing-else/post/5421938/#5421938

[@maikdean8490]

What’s the maximum megapixel camera it can handle please reply thanks 🙏

[@akasitchantanit4362]

what is Model of camera for test in this lab ?

[@내아이디-n5n]

Thank you for teaching but i have a problem AttributeError: 'jetson.utils.gstCamera' object has no attribute 'CaputureRGBA' could you tell me what is the problem??

[@dusty-nv]

Hi, I think you made a typo - it should be 'CaptureRGBA' instead of 'CaputureRGBA'

[@feelgood413]

Does this code work in NVIDIA PX2? Do we have to make some modification ?

[@dusty-nv]

I haven't tested it on DRIVE PX2, however perhaps you should be able to compile it. It does use GStreamer for the CSI camera, and I'm not sure if that's provided on DRIVE platform or not.

[@dowiee2694]

That dog is so fucking cute!!!

[@arisatrombley8140]

Can you use these object detection models within ISAAC ROS on GPU? jetson inference builds within the isaac ros container but it doesn't have the pop up for me to select which model to download. I wanted to use the people net model

[@arisatrombley8140]

im using a jetson orin nano with cuda 11.4

[@grumpyyyrider]

for anyone who got the gst error with RGBA, use "0" instead of "/dev/video0"

[@gprashant840]

Dude you are awsome!!!

[@ashok_ign5623]

PowerServiceHwVic::cleanupResources i got this after first frame and code stopped please help ??

[@feelgood413]

How to use similar code for NVIDIA Px2?

[@toordog1753]

Yeah, that's not object detection, that's special detection...

[@뉴식]

This cording can use GPU? and How can I train dataset ?

[@laurencevanhelsuwe3052]

All these needed steps are just insanely user unfriendly. Think about that and consider what impact this has on sales. Look at how Tinkerforge (the DE company) manages to bring technical stuff to its users without burying them in accidental complexity.

[@AlanWagoner]

Damn that's cool tech!

[@amirarjmand5295]

does have NVIDIA any industrial gpu?

[@Sajanrajtd]

excellent......

[@romangerasimov3523]

hmmmm... if it add it to drone... and write to increase speed to contact, and center image at center of screen (by servos) = hunter-seeker terminator ps: waiting for 1st terminator algoritm (I'm too lazy to write)

[@julayelei]

Cool!

[@gwitichis1]

CNN or NN object detection needs high power computing and is a dead end for small/mobile applications. Object detection with the IBM true north chip is low power due to spike technology.

[@stahelpeter]

Very interesting tutorial. I run the python code from either Mobaxterm or VNC desktop but always get this error [OpenGL] failed to create X11 Window. (have a Logitec C920 connected at /dev/video0) Any hint what the problem could be? Does anyone have the demo running with my setup? (no monitor on HDMI)

[@fernandohenriques6754]

You need to activate xforward! In mobaxterm u just need to click in the button X on upper right corner that says "X server"!

[@stahelpeter]

@@fernandohenriques6754 Thank you for the answer. After a new installation I'm able to run python scripts like imagenet-camera.py with USB or CSI camera within VNC without problems. When I run the same imagenet-camera.py in mobaxterm (X11 server running) a black window opens and after a while it ends with short dump. To open camera windows in mobaxterm with OpenCV works fine. Can you confirm you are able to run python scripts from the tutorial within mobaxterm without problems? For me it looks like the issue is the methode in jetson.utils package (camera = jetson.utils.gstCamera(opt.width, opt.height, opt.camera))

[@Iammrbt2]

Doesn't work with an RTSP data stream :/

[@joelguino-o4916]

thank you..

[@SunilKumar-wh6ph]

How to install Jetson nano in a laptop?

[@rudreshmehta6510]

I do not have jetson , any other way?

[@robotics4.077]

Thanks :D

[@giangonzalez3283]

Can you use the same code on a raspberry?

[@FalguniDasShuvo]

It would have been nicer if it was detecting itself ( a Jetson Nano ) [ 22:54 ]

[@apzzbn]

he looks like Roy from "The Office"

[@tunaakyol2579]

how can i rotate the camera ? I am using detectnet.py

[@dusty-nv]

You can change the flip-method here: github.com/dusty-nv/jetson-utils/blob/798c416c175d509571859c9290257bd5cce1fd63/camera/gstCamera.cpp#L416 Then re-run 'make' and 'sudo make install'. You can find the values for flip-method by running 'gst-inspect-1.0 nvarguscamerasrc'

[@saibollu7073]

Does anyone konow How to get the exact coordinates of the perticular object in frame and area of the object

[@Goodwill345]

You need to get the pose of the object, pose means x,y,z and rotations total 6dof. You need to transpose camera coordinates to world coordinates. There must be a DNN for this already

[@dusty-nv]

You can directly access the members of the detection results, including coordinates, area, classID, confidence, ect. See the documentation under 'jetson.inference.detectNet.Detection' here: rawgit.com/dusty-nv/jetson-inference/python/docs/html/python/jetson.inference.html#detectNet

[@ashwinkp7828]

hey i have a doubt, did you any algorithm in this

[@adilhossain227]

Sir, i have my custom trained inference graph (based on SSD inception), how can I run that in this code?

[@payloan4558]

Is that running on android?

[@dusty-nv]

It's running on Ubuntu from NVIDIA JetPack - developer.nvidia.com/embedded/jetpack

[@PranshuTople]

How to run similar object detection on a .MP4 video file that I have in the same directory

[@dusty-nv]

Hi Pranshu, see this post to run it on video: devtalk.nvidia.com/default/topic/1071209/jetson-nano/detectnet-video-/post/5427509/#5427509 You may need to convert your MP4 file using ffmpeg tool or similar, to a format/codec that OpenCV is able to read.

[@OthmaneGUESSOUS]

il est ou detect net console

[@andyrodrigoalvarado118]

I wish more time, why I'm so busy :(

[@andrzejostrowski5579]

Am I the only person bothered by this API not following PEP-8 rules?

[@MrBratkenSolov]

let me guess, you call library with 200k lines of code in your project with 10 lines of code? Stupid clickbaits **checks video** yep

[@enarm5947]

after running the script, my nano power off

[@dusty-nv]

Before running the script, try setting the power mode to 5W: $ sudo nvpmodel -m 1 If your Nano no longer powers off in 5W mode, it may be that your power supply doesn't deliver a consistent 10W. See this thread for some other power supplies that you could try: devtalk.nvidia.com/default/topic/1048640/jetson-nano/power-supply-considerations-for-jetson-nano-developer-kit/

[@enarm5947]

@@dusty-nv thank you, worked

[@ewaldsteven]

Please don't kill me.

[@Praxss]

Nvidia with linux....

[@leeornehardea7137]

I get this error: jetson.inference.__init__.py Traceback (most recent call last): File "./detectnet-console.py", line 24, in import jetson.inference File "/usr/lib/python3.6/dist-packages/jetson/inference/__init__.py", line 4, in from jetson_inference_python import * ImportError: dynamic module does not define init function (initjetson_inference_python) I installed everything and followed the Github page but no luck. Any ideas?

[@fvw94]

I have a similar problem, still figuring out how to fix it ImportError: No module named jetson.inference

[@dusty-nv]

@@fvw94 if you run the program as "python3 detectnet-console.py" or "python detectnet-console.py", are you able to get past the import?

[@fvw94]

@@dusty-nv yep, I needed to add python3

[@leeornehardea7137]

@@fvw94 did it work for you that way?

[@fvw94]

@@leeornehardea7137 yes it did :D

[@bigeye6525]

NB!

[@leonc2940]

听力时间

[@skewty]

how typical.. Released with an old outdated version of Ubuntu. That should be a clear indication to everybody that NVIDIA doesn't really care to support this well. I'd caution investing too much into development on this platform especially since NVIDIA is notorious for being a terrible linux supporter.

[@khatharrmalkavian3306]

10 lines of code? Why don't you just call it one line on the command-line? What a quack.

[@gachihard7397]

Absolute lifesaver thanks king 🫡