yolov8n runs ~100x slower than reported in model-zoo

[talcs]

Hello,

In the Hailo8 model zoo, yolov8n is listed with FPS(batch size=1) of 1036 frames.

However, when compiling the Hailo8 C++ object detection example and running it with yolov8n on the provided video full_mov_slow.mp4, it works pretty slow and it finally reports FPS=11.076 and latency=90.2852ms.

In the past there were multiple object detection examples (sync and async, maybe more...) and I could run the models much faster than now.

How can I accelerate the current object detection example?

[majedabu]

Hi

The reported 11 FPS is not accurate — when saving video, the app fixes playback at ~30 FPS for smooth viewing.
In my tests with yolov8n on full_mov_slow.mp4, the actual throughput was ~44 FPS, not 11.

Currently, the example always saves the processed video at 30 FPS, which explains the lower number.

In the next release we’ve added an option to choose between saving or displaying.
Using the new -s flag will save the processed video (capped at ~30 FPS).
Running without -s will be the new default: display mode with performance above ~230 FPS.