Hello, I'm doing some work with an OAK-D PoE using a custom-trained YoloSpatialDetectionNetwork. I'm trying to optimise for high FPS and low latency, and I'm looking for benchmarks to compare the performance of my code against. There is a NN performance table [here](https://docs.luxonis.com/projects/hardware/en/latest/pages/rvc/rvc2/#rvc2) and it links to more detailed info [here](https://docs.google.com/spreadsheets/d/1yMD4L3gNTkv9d-CqwHTYkn1_on9qDwjeDSUiW2x_H8k/edit#gid=0 "detailed table"). However these figures are explicitly for USB3, and I *assume* they're for YoloDetectionNetwork (not Yolo**Spatial**DetectionNetwork) based on the number of SHAVES used. * Is there any equivalent data for PoE? * Is it possible to get copies of the scripts that were used to generate these benchmarks so I can try it on my own hardware with my custom model? * Are there any equivalent NN performance figures using YoloSpatialDetectionNetwork? Thanks very much

Did you see this other chart? * https://docs.google.com/spreadsheets/u/8/d/e/2PACX-1vQ_tVk2PhOhnFeJrL5t2rtncxHeDVYX8j1o52vdZozRzXJ5C3gq8EngVvx32suCPasIelIwU5Ny6tLE/pubhtml?gid=41416082&single=true It gives much more detailed info, included the source.

Hi @"bryan1"#p10834 , 1. We don't, but from pure AI inference speed, USB and POE should be equivalent. There are other factors, like higher CPU usage on POE, as it has to handle the networking stack. What's the discrepancy of inference FPS between USB and POE you have noticed? 2. We've used this script: https://gist.github.com/Erol444/8cf8c546bb5255253feb94e3b55814d9 3. I don't think so, but it requires quite a lot more processing, but I believe it shouldn't be a bottleneck (additional stereo calculations and NN decoding). Perhaps you could share the model + minimal repro code, and we can try it out locally as well. Thoughts?

@"erik"#p10869 Thanks for the detailed response. 1. I'll try to log some data and get back to you. 2. Thanks very much, this is exactly the sort of script I was hoping you would have. 3. I'll get a minimal script together.

NN Performance Benchmarks for OAK-D PoE and YoloSpatialDetectionNetwork

bryan1

Hello,

I'm doing some work with an OAK-D PoE using a custom-trained YoloSpatialDetectionNetwork. I'm trying to optimise for high FPS and low latency, and I'm looking for benchmarks to compare the performance of my code against.

There is a NN performance table here and it links to more detailed info here. However these figures are explicitly for USB3, and I assume they're for YoloDetectionNetwork (not YoloSpatialDetectionNetwork) based on the number of SHAVES used.

Is there any equivalent data for PoE?
Is it possible to get copies of the scripts that were used to generate these benchmarks so I can try it on my own hardware with my custom model?
Are there any equivalent NN performance figures using YoloSpatialDetectionNetwork?

Thanks very much

vital

Did you see this other chart?

It gives much more detailed info, included the source.

bryan1

vital Thanks, but unfortunately that chart only has object detection (green) for yolo, not depth estimation (purple) and it's still USB3. And as best as I can tell, the linked sources are for the models; to be more clear, I'm interested in the pipeline script that loads and runs the models.

erik

Hi bryan1 ,

We don't, but from pure AI inference speed, USB and POE should be equivalent. There are other factors, like higher CPU usage on POE, as it has to handle the networking stack. What's the discrepancy of inference FPS between USB and POE you have noticed?
We've used this script:
I don't think so, but it requires quite a lot more processing, but I believe it shouldn't be a bottleneck (additional stereo calculations and NN decoding). Perhaps you could share the model + minimal repro code, and we can try it out locally as well.

Thoughts?

bryan1

erik

Thanks for the detailed response.

I'll try to log some data and get back to you.
Thanks very much, this is exactly the sort of script I was hoping you would have.
I'll get a minimal script together.