Bad performance of jetson-inference with ssd-mobilenet-v2

I’m just starting out with the Jetson Nano (4G RAM) and I wanted to try out object detection.

I cloned the jetson-interence GitHub-Repository and followed the installation instructions.

Then I ran ./detectnet-console --network=ssd-mobilenet-v2 images/peds_0.jpg images/test/output.jpg to try it out.

However it’s performance was significantly worse than in the demo video :

[TRT]    ------------------------------------------------
[TRT]    Timing Report networks/SSD-Mobilenet-v2/ssd_mobilenet_v2_coco.uff
[TRT]    ------------------------------------------------
[TRT]    Pre-Process   CPU   0.07917ms  CUDA   1.00812ms
[TRT]    Network       CPU 802.01288ms  CUDA 800.92664ms
[TRT]    Post-Process  CPU   0.09828ms  CUDA   0.09844ms
[TRT]    Visualize     CPU  40.18549ms  CUDA  40.77422ms
[TRT]    Total         CPU 842.37579ms  CUDA 842.80743ms
[TRT]    ------------------------------------------------

Note: I executed sudo jetson_clocks and sudo ldconfig before this and the NV Power Mode is set to 0 (MAXN)

Some additional information may also be the following error printed during the execution:
[TRT] Could not register plugin creator - ::FlattenConcat_TRT version 1

As well as the following warning given:
[TRT] Using an engine plan file across different models of devices is not recommended and is likely to affect performance or even cause errors.

Any ideas why this is would be appreciated, as I already looked at most of the related posts in this forum and I’m really a beginner to this.

Hi @moritz.ahrens2, running on a single image one time will result in variable performance and isn’t recommended for benchmarking. In particular, the first run will often take longer than normal as the CUDA kernels are loaded for the first execution.

To get better performance information, it is recommended to process a continuous live camera stream or video. You can see how to run on a video from disk in this step: