Hi,
Please check my reply here.
In summary, you can expect to see a growing latency if nvinfer takes longer to process that the time of one frame. There are several ways to optimize your pipeline in order to avoid this:
- Adding
is-live=trueto your source - Adding
sync=falseto your sink - Adding leaky queues before the bottleneck elements:
queue leaky=2 max-size-buffers=1 - Set
interval=ninnvinferto skip inference in some buffers