• Hardware Platform: GPU
• DeepStream Version: 7.1
• TensorRT Version: 10.3.0.26
• NVIDIA GPU Driver Version: 535.183.01
• Issue Type: Question
In my pipeline, I would like to do inference every 5 frames and use the NvDCF tracker’s past-frame data to display bboxes even on frames where inference was not running.
From a few discussions here I understand that building a custom plugin to buffer frames in advance is necessary:
- How to Reduce BBoxes Flickering on Deepstream
- Sustain bounding boxes in DS 6.0 - #27 by kesong
- Sustain bbox by tracker in DeepStream 5.1
I made my own plugin and I do manage to fetch the past-frame data to display the bboxes that did not have detections (because inference was not running)
However, I am running into a strange issue:
In my custom plugin, If I buffer 1, 2 or 3 frames (and set a proper inference interval) then my plugin works and I get the desirred result of smooth bboxes!
If I buffer 4 frames the pipeline completly stalls!
Here is a minimal working example of the plugin to reproduce the issue (I removed the past-frame fetching logic as i have no issue with it, so this plugin just buffers frames until it has enough and then releases them):
frame_buffering.txt (3.7 KB)
Do you have any idea why this plugin would stop working for a min_delay > 3?
Do you confirm that this is the preferred way to achieve the goal of no flickering bboxes? Is there a better way?
Thank you for reading