Hello,
i have successfully installed CUPY on the Jetson Xavier NX.
Very easy to install (many thanks kmaehashi for your help) :
https://docs.cupy.dev/en/latest/install.html#installing-cupy
If you have JetPack 4.X.X :
pip install cupy-cuda102 -f https://pip.cupy.dev/aarch64
If you have JetPack 5.0.2 :
pip install cupy-cuda11x -f https://pip.cupy.dev/aarch64
CUPY is really interesting it i guess it is worth giving it a try.
Alain
Hello Honey-Patouceul,
i tried your code with the Jetson Xvier NX and it works. Just a little problem with colors (red and Blue channels are inverted).
Concerning speed, i am not really sure it is faster than with opencv. I will see that.
Alain
Hi Alain,
You may tell if this issue is seen with HQ (raw) format, encoded (H264) or both.
I did remove a RGB2BGR conversion that I felt not needed when using JetsonTreatement, but if the pushed frames are in RGB format rather than BGR, you would try to adjust such as:
gst_pipe = "appsrc ! video/x-raw,format=RGB,width=...
Hello,
the issue is for both format. I will try your proposal.
I made a test with vcpkg.exe with my laptop. Interesting. Unfortunately, it seems there is an issue with CUDA 11.6. I will give it a another try later.
I also tried to compile opencv with classical method (Visual Studio) but i get some compile errors. Not a big surprise for me. It is quite funny to try to compile a big library. 3 hours to get errors. When i will have time, i will make a clean reinstall on my laptop. I guess i have something not really clean in the laptop.
Alain
Hello everyone,
i quick message to wish you a merry Christmas !
Many thanks to people who participate to this topic.
Special thanks to Dusty_nv !
Alain
Have a good christmas time too and take care of yourself!
Happy new year !
Alain
Hello,
this year, i will get back with Moon and planetary imaging. My Celestron C9.25 is sleeping for too much time.
I will still work on colours of the Moon and i also plan to test my adaptative absorber noise removal filter with atmospheric turbulences.
As Moon and planetary imaging need very high frame rate, i will have to find the way to get a more powerful laptop with as many as possible CUDA cores. Jensen, if you read me, i would like to talk with you.
Alain
I have tried out chatgpt which gives me this alternatives to the code-line for video-writing. Maybe it is worth to look at:
Here are a few alternative ways to write the line of code you provided:
opencv-python-headless library:Copy code
video = cv2.VideoWriter(os.path.join(video_path,nom_video), cv2.CAP_FFMPEG, 25, (width, height), isColor = True)
moviepy library:Copy code
from moviepy.video.io.VideoFileClip import VideoFileClip video = VideoFileClip(os.path.join(video_path,nom_video)).fps(25).resize((width,height)).write_videofile(nom_video)
pyav library:Copy code
import av video = av.open(os.path.join(video_path,nom_video), 'w') stream = video.add_stream('vp9', 25) stream.width = width stream.height = height
Please keep in mind that, depending on the library and the specific use case, you might have to adjust the parameters and methods used in these alternatives.
One last try:
Here are a few alternative ways to write the line of code you provided using OpenCV:
cv2.CAP_FFMPEG instead of fourcc:Copy code
video = cv2.VideoWriter(os.path.join(video_path,nom_video), cv2.CAP_FFMPEG, 25, (width, height), isColor = True)
cv2.CAP_GSTREAMER instead of fourcc:Copy code
video = cv2.VideoWriter(os.path.join(video_path,nom_video), cv2.CAP_GSTREAMER, 25, (width, height), isColor = True)
cv2.CAP_AVFOUNDATION instead of fourcc on macOS:Copy code
video = cv2.VideoWriter(os.path.join(video_path,nom_video), cv2.CAP_AVFOUNDATION, 25, (width, height), isColor = True)
cv2.CAP_DSHOW instead of fourcc on Windows:Copy code
video = cv2.VideoWriter(os.path.join(video_path,nom_video), cv2.CAP_DSHOW, 25, (width, height), isColor = True)
cv2.CAP_MSMF instead of fourcc on Windows:Copy code
video = cv2.VideoWriter(os.path.join(video_path,nom_video), cv2.CAP_MSMF, 25, (width, height), isColor = True)
Itās important to note that, depending on the
Donāt know if that will help and yes I know, it is not as easy to use this chat bot and then all problems are gone. It is only a try for finding some (new) ideas.
I doubt that chatGPT has ever tried what it proposes ;-)
The generic call of opencv VideoWriter would be:
cv2.VideoWriter(desc_str, VideoBackend, 4CC, float(framerate), (int(w),int(h)), bool(isColor))
where VideoBackend is the backend API, such as cv2.CAP_FFMPEG, cv2.CAP_GSTREAMERā¦
If the backend is FFMPEG, then the container type would be guessed from the file extension of the filepath given as desc_str, and the encoding would be according to the 4CC, such as:
# This would encode with X264 codec and mux into avi container
cv2.VideoWriter('Test.avi', cv2.CAP_FFMPEG, cv2.VideoWriter_fourcc(*'X264'), float(25), ...
If the backend is GSTREAMER, the desc_str is the pipeline from appsrc to the wanted sink, and CC4 is 0 (RAW):
# This would encode into H264 with HW NVENC on Jetson and mux into an AVI container:
cv2.VideoWriter('appsrc ! queue ! video/x-raw,format=BGR ! videoconvert ! video/x-raw,format=BGRx ! nvvidconv ! video/x-raw(memory:NVMM,format=NV12 ! nvv4l2h264enc ! h264parse ! avimux ! filesink location=Test.avi', cv2.CAP_GSTREAMER, 0, float(25), ...
Hope this will help chatGPT or else for next versions ;-)
LOL, ok it was a try ;-)))))))))))))))))
Hi guys,
ChatGPT will replace us all soon ! Or not.
Alain
ChatGPT still has a lot of training left. See:
https://www.tomshardware.com/news/chatgpt-told-me-break-my-cpu
Skynet is a bit young but he will be back stronger and smarter!
Yesterday, a quick deep sky survey, using AGX Orin and JetsonSky :
Many satellites.
I added HDR capture in a beta version of JetsonSky. I need to test it on easy targets like M42 (Orion Nebula, very hard to get without HDR because of the brightness of the nebula). Need some time to make such tests.
Alain
Hello,
last night test with JetsonSky and the AGX Orin. The main target was Orion nebula (M42).
I used 50mm F1.4 and 135mm F2.5 lens.
The result is not so bad but not really good. The 135mm gives more things but F2.5 is too much. My kingdom for a Canon 85mm F1.2 !
HDR did not give expected results. But as the mount was not calibrated, there were no earth rotation compensation and as the exposure time was a bit long (from 500ms to 1s), i guess the HDR routine could not work properly. Sometimes it worked and the result was interesting with M42 but most of time, it was an epic fail.
Still some work to do.
Alain
Hello,
I was thinking about further really improvements for my project and i need to consider to get better equipment (lens, cameras, laptop computer with many CUDA cores) to get significant better results.
I am not sure it is the right place to ask for help and if it is not, i apologize and i will delete this post.
Well, i do need better equipment but as my equipment is not that bad, it means i need expensive equipment. For 8 years, i bought myself my equipment but this time, the level is a bit high.
So, a friend of mine told me to make a funding pot. So i did.
I donāt know if you can participate from an other country than France.
So, if you find my work interesting and you want to help me getting better equipment, you can go there :
It is all in French and i am very sorry for that.
The fund target is very high so donāt be surprised. I try to get very good equipment and very good equipment is always very expensive.
You can also share the funding pot if you want. This will allow me to touch more people.
Many thanks in advance.
Alain
It seems I had a bad idea. Sometime, we choose wrong solution to solve a problem.
Anyway, I continue my work but I reduce support. Just focusing on my needs.
So, forget previous post. You can post here freely without fear, I wonāt ask you for money, I swear š
Clear sky.
Alain
To not forget :
pip install pycuda --user
sudo ln -s /usr/include/locale.h /usr/include/xlocale.h