I’m working on a real-time computer vision project. I’m currently using a Jetson TX1 board with the latest jetpack and a USB OV1064O camera (UVC compliant) delivering 1280x1080@30. Recently, I’ve started measuring latency with different pipelines. The way I measure latency is the following:
- Print a timestamp in milliseconds since the epoch before capturing a frame.
- Call the function for capturing a frame.
- Draw that timestamp over the image.
Then, I aim the camera to my monitor to get the timestamps being printed and afterwards I compare the recorded timestamp with the one drawn over the image.
My results are:
CODE FRAME TIME (ms) LATENCY (ms)
MIN MAX AVG MIN MAX AVG
OpenCV 17 44 25.3 130 198 168.4
gst-camera (jetson-inference) 19 27 21.7 99 114 105.2
v4l2cuda (tegra_multimedia_api) 27 37 30.1 84 116 106.5
I find it hard to believe that for most real-time applications 100ms (around 3 frames at 30fps) is acceptable.
Does this benchmark sound plausible to you?
Is there any way I could improve latency with this setup?
Thanks in advance.
For going through UVC driver, you may try ‘sudo ~/jetson_clocks.sh’ to get better performance because the buffers are CPU buffers.
The reported results were taken after having run the jetson_clocks.sh script.
We try with See3CAM_CU135 and see 9-frame latency(300ms) on r28.1(Kernel 4.4):
nvidia@tegra-ubuntu:~/tegra_multimedia_api/samples/12_camera_v4l2_cuda$ ./camera_v4l2_cuda -d /dev/video0 -s 1920x1080 -f UYVY
Failed to query video capabilities: Inappropriate ioctl for device
libv4l2_nvvidconv (0):(774) (INFO) : Allocating (4) OUTPUT PLANE BUFFERS Layout=0
libv4l2_nvvidconv (0):(790) (INFO) : Allocating (4) CAPTURE PLANE BUFFERS Layout=0
[INFO] (NvEglRenderer.cpp:109) <renderer0> Setting Screen width 1920 height 1080
^CQuit due to exit command from user!
----------- Element = renderer0 -----------
Total Profiling time = 319.323
Average FPS = 30.0009
Total units processed = 9581
Num. of late units = 1188
The result may vary with different USB cameras. The UVC video driver is default in K4.4. Not sure but there might be room for improvement. Other users may share experience.