Camera Capture YUV: Jetson-utils

Autonomous Machines Jetson & Embedded Systems Jetson Xavier NX
1

Hi,

I will work in future with a YUV Camera and my aim is to display the pictures in a fast way and use opencv functions.
I already used the Capture function of jetson-inference for decode mp4-videos. My question is now concerning this link jetson-inference/aux-image.md at master · dusty-nv/jetson-inference · GitHub under “IMAGE FORMAT”. There it is said that a “variety of other format are defined for image acquisition” including YUV fomats.
But also the Capture function of jetson-utils just accepts RGB formats.
That is my question. Can the Capture function can be fixed that it accepts YUV formats?
If NOT: What would be another approach to reach my aim? If anyone is experienced with that, I am very glad to get more information of how fixing this.

The question is similar to this one - How to convert yuv to rgb8/rgb8a/rgb32/rgb32f to use Imagenet/detectnet/segnet in jetson-inference - but I can not take a solution from the answer. Therefor please give me more detailed help.
I am looking forward to your help!

Best greetings, J. W.

2

I suggest asking this question on one of the Jetson forums.

1 Like
3

Hi @user152011, jetson-utils should handle cameras that transmit video in YUV formats, and typically it converts these to an RGB/RGBA format in the Capture() function. Are you saying that you want the YUV data returned from Capture() ?

I haven’t really tested that use-case, but I traced the code again and it looks like you can request any IMAGE_FORMAT that the cudaConvertColor() function supports the conversion of: jetson-utils/gstBufferManager.cpp at 8b6c5ca4b2b52d51f5415ce033ee52563ddf3f2c · dusty-nv/jetson-utils · GitHub

Seeing as cudaConvertColor() doesn’t support converting YUV formats to other YUV formats (only YUV->RGB/RGBA), I would request the original YUV format the pipeline uses such that no colorspace conversion is required.

1 Like
5

Hi dusty_nv!

Thank you for your answer! First of all it seems to be good news! But I have to try! I will reply here if the topic will be still open (It could take more than 14 days).

  • Yes. So the input is YUV 4:2:2 and in the end I would like to have to have a YUV format that can be split in 3 channels by OpenCV. But if I understand your last sentence correctly, it is not possible to convert YUV 4:2:2 into such a more handy YUV format, right?

Does it mean that I have to comment the marked line 435 and instead I return the latestYUV which has then the YUV 4:2:2 format? But then I don’t receive YUV in 3 channels, do I?

One question beyond that: Why is the NV12 format used by the Capture function? For me it seems like a temporary format as well the input format is RGB or YUV. The command to use NV12 is given by the gstreamer pipeline https://github.com/dusty-nv/jetson-utils/blob/8b6c5ca4b2b52d51f5415ce033ee52563ddf3f2c/camera/gstCamera.cpp#L150 , is that correct?

Thank you again! I am looking forward to your answer!

6

It appears in the source of cudaConvertColor(), there is RGB/RGBA → I420 and RGB/RGBA → YV12 implemented. There is no cross-YUV conversion, but you could request RGB8 for example from Capture(), then use cudaConvertColor() to convert it to I420/YV12 if those are friendlier formats for you.

The CSI and hw decoder elements use NV12 so it skips the temporary format conversion by reading it directly.

8

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.