Is DriveNet only support RAW image?


The DriveNet is use video RAW data to do inference.
Could we use the h.264 video file to do this? Because the size of video RAW data is too big. Any suggestion?


Hello KentTeng,

Unfortunately DriveNet can’t support h.264 video file, so please use RAW video data for DriveNet. Thanks.

I have saome h.264 video file, how can I convert to RAW video, any tool?


I have saome h.264 video file, how can I convert to RAW video, any tool?


Dear liukun,

You may compile ffmpeg from sources (compile with --enable-pic for position independent code) on your drive pix. Then, you may convert from raw to x264, and back. Please refer to below link. Thanks.

Hello, I would like to do the same thing as liunkun.

I have tried these steps and I am able to convert a h264 file into raw video. However, when I try to play that video file with DriveNet, I get this error:

“DW_FILE_INVALID: CameraVirtual: video file has wrong magic number”

I cannot find any documentation in driveworks about what this exception means.


Hi KentTeng, liukun and Matthew_vsi_labs,

It is possible to feed h.264 videos as input for DriveNet, see this thread for more information:
[url]Object detection using driveworks - DriveWorks - NVIDIA Developer Forums
There might be a loss of accuracy on the object detection, but depending on what you are working on, this might be acceptable.

Hey Frank_cisco,
That’s the object detection sample you mentioned. DriveNet is preferred because it can detect multiple objects at same time by default. (from same frame). Any way, it seems there is no way to feed anything except .RAW video to DriveNet! I’m interested to receive advise on it!

Hi rghahremani,

DriveNet can also use H264 videos as input. You just need to execute it like this:
sample_drivenet --video=<pat/to/file/filename>.h264.

Like I mentioned before, there is a loss of accuracy when using this format instead of RAW.