Replaying the same fifteen-second clip on every channel makes the load reproducible, but it can also synchronize the expensive parts of the workload. For the helmet classifier, all eight streams may reach their busiest target count at the same time.
I would repeat the staircase with phase offsets between streams and with a deliberately synchronized crowd burst, retaining crop counts alongside the FPS floor. Those two runs would show whether capacity depends on aggregate average demand or short classifier bursts. That complements your distinction between task-frame updates and accelerator invocations without treating eight passing channels as a universal hardware limit.
Good point. Reusing the same clip controls the scene content, but stream timing could still change how much classifier work arrives at once. We haven’t measured that alignment, so I wouldn’t assume the streams were either synchronized or independent.
Comparing staggered playback with a synchronized crowd burst would help separate average demand from burst pressure. I’d track the crops reaching the classifier and queue depth alongside the FPS floor, to see whether similar average rates hide different backlogs.
That would be a useful extension of this benchmark. Thanks for calling it out.
Agreed, the alignment needs measuring rather than inferring it from the shared clip. Queue depth alongside the FPS floor would make that distinction much clearer.
I would add the age of the frame when its result is emitted, because a system can maintain its reported processing rate while falling behind the live scene. Comparing that age and the time taken to drain the backlog after each burst would show whether staggered and synchronized playback differ mainly in temporary delay or in sustained capacity.
Ahmet Özel
AI Engineer. Computer Vision, RAG and LLM agents.
Replaying the same fifteen-second clip on every channel makes the load reproducible, but it can also synchronize the expensive parts of the workload. For the helmet classifier, all eight streams may reach their busiest target count at the same time.
I would repeat the staircase with phase offsets between streams and with a deliberately synchronized crowd burst, retaining crop counts alongside the FPS floor. Those two runs would show whether capacity depends on aggregate average demand or short classifier bursts. That complements your distinction between task-frame updates and accelerator invocations without treating eight passing channels as a universal hardware limit.