Deepstream OSD flickering

Accelerated Computing Intelligent Video Analytics DeepStream SDK
#1

Please provide complete information as applicable to your setup.

• Hardware Platform: Jetson
• DeepStream Version 5.0
• TensorRT Version 7.1
• Issue Type: bug

I’m running yolov4 in the deepstream pipeline on the Jetson AGX. In the resulting video with detections bounded by a box, there’s a flickering bounding box border, where in any particular image, one or more of the box borders are grey instead of red (the configured color). The video is 2448x2048 running at 22fps, and I use the deepstream’s streammux to reduce it to 608x508 for inference.

I tried turning on and off the tracker, which does not affect this bug. I also tried adjusted the primary-gie interval to 0 and other values and the flicker is still there. Please find a video of the flickering here

For reference, here is my deepstream_app_config_yoloV4.txt, as well as my tracker_config.ymldeepstream_app_config_yoloV4.txt (4.4 KB) tracker_config.yml (6.6 KB)

#3

Hey, which model are you using?

#4

Might be cause by H264 compression.
Try with H265 or a higher bitrate.
They color flickering should also disappear if you use EglSink instead of File sink.

#5

Thank you for the responses!
@Blard.Theophile, I tried the H265 encoder, and the H264 encoder at 55 Mbit/s (sufficient for 608x508 pixel image at 22 fps), and the flickering remains. Note that I am also using the NVENC encoder engine.

@bcao, I am using the yolov4 model, with the NvDCF tracker and OSD enabled. The problem persists with the yolov4 model even when I turn off the tracker. Note that I did not see this problem with the yolov3 model + tracker + OSD.

I suspect it has something to do with the yolov4 model. I confirmed that the bounding box coordinates out of the yolov4 model are present for every frame. I’m not sure where else to look to debug this issue.

Thank you again for the help!

#6

Should I file a bug report about this issue?

#7

Hey Julie,
I mean which yolov4 model are you using, can you confirm the model’s output is well without DS pipeline?
Nvidia also provide an official deepstream yolov4 sample https://github.com/NVIDIA-AI-IOT/yolov4_deepstream