• OAK-D-ProPOE camera FPS is slow

5 days later

Hi jakaskerl,

I have two different OAK devices: an OAK-D USB3 and an OAK-D Pro POE. When I run my pipeline with the first device, I receive 26 FPS. However, when I run it with the second device, the frame rate drops to less than 10 FPS. I am using yolov6n-coco640x640 and ObjectTracker to track and count people.

I have connected my OAK-D POE to my PC using a Cudy 30W Gigabit POE/POE+ Injector model POE200, which supports up to 1000 Mbps Ethernet, and CAT 7 cable. I connected my OAK-D with a USB3 cable.

When I run poe_test.py, I receive good performance, like in the image below:

Also I tried to run this simple code with both devices,

from depthai_sdk import OakCamera
with OakCamera() as oak:
color = oak.create_camera('color')
oak.visualize(color.out.main, fps=True)
oak.start(blocking=True)

The OAK-D USB3 yielded an average of 30 FPS, while the OAK-D Pro POE only achieved around 3 FPS.

I'm not sure how to proceed or what steps I can take to resolve these issues. I would greatly appreciate any tips or suggestions you can provide.

Thank you.

      Hi Irena
      The PC. Looks like your RJ45 connector on your motherboard only supports up to 100Mbps bandwidth. Can you confirm this is the case? You could benefit from a secondary PCIE network adapter if you are using a desktop PC.

      Thanks,
      Jaka

          Hi jakaskerl

          I use a CAT 7 RJ45 cable (600MHz) witch is up to 10 Gbps.

          Could you provide me with some detailed information about these adapters?

          I am seeking alternative methods to verify the configuration of my POE-cam. Are there any other tools or techniques you would recommend. I'm not sure what about this discrepancy in the FPS speed between the Oak-D and Oak-D POE cams

          Thanks a lot!

              Hi Irena
              The UTP cable is more than capable of carrying speeds needed for optimal depthai communication. The problem is (probably; this is what I wish to know) the motherboard of your computer. The motherboard has a chip specifically designed to process network traffic. If the chip only supports up to 100Mbps bandwidth, then this will create a bottleneck for the whole network pipeline.

              For example, I'm running a 10y/o laptop which only supports speeds up to 100M. I have it connected to the poe switch (1000M) using cat5e cable (1000M). The most I can get out of the network is 100M as my laptop is the limiting factor.

              You can buy a network adapter which can be connected to your PCI-e slot on your motherboard (same type slot as the graphics card) if you own a desktop PC --> means your PC is modular and upgradeable.

              FPS - bandwidth calculations: https://docs.luxonis.com/projects/api/en/latest/tutorials/low-latency/#bandwidth

              Can you check what network speeds your motherboard supports?

              Thanks,
              Jaka

                Hello @jakaskerl,

                Thank you for your prompt reply.

                You are correct, my network speed is 100 Mbps.

                Therefore, I would like to inquire if you believe that this is the cause for obtaining only 3 FPS when running the viewer code (specifically for viewing the colour camera) with the POE cam. I still have doubts regarding whether this issue is also related to a bottleneck.

                Thank you once again for your assistance.

                  Hi Irena
                  I only suggested bandwidth since it's the most probable issue, but it really depends on the amount of data you are trying to send to your host machine. You can either try to calculate the bandwidth requirements using above link or you can try lowering the resolution of the frames you are trying to send through. I believe that since you stated

                  Irena (specifically for viewing the colour camera)

                  it should probably be bandwidth since color at same resolution uses 3 times as much bandwidth as mono (RGB vs G).

                  Can you try using cam_test.py to ensure it is not the viewer that's causing low fps?

                  Thanks,
                  Jaka

                      Hi @jakaskerl

                      I just sent you the config pram when I run cam_test.py...not sure if is it ok.

                      After connect with OAK-D PRO POE cam, here the results:

                      Connected cameras:
 -socket RGB   : IMX378 4056 x 3040 focus:auto  - COLOR
 -socket LEFT  : OV9282 1280 x  800 focus:fixed - MONO
 -socket RIGHT : OV9282 1280 x  800 focus:fixed - MONO
USB speed: UNKNOWN
IR drivers: [('LM3644', 2, 99)]
Cam:      rgb           left          right         camd     [host | capture timestamp]
FPS:   2.24|  2.81   2.23|  2.89   2.25|  2.90   0.00|  0.00 [19443010C1C3741300] [169.254.1.222] [9.114] [ColorCamera(8)] [error] Camera not detected on socket: 3

                      This are OKA-D cam results:

                      Connected cameras:
 -socket RGB   : IMX378 4056 x 3040 focus:auto  - COLOR
 -socket LEFT  : OV9282 1280 x  800 focus:fixed - MONO
 -socket RIGHT : OV9282 1280 x  800 focus:fixed - MONO
USB speed: SUPER
IR drivers: []
Cam:      rgb           left          right         camd     [host | capture timestamp]
FPS:  14.22| 14.29  14.12| 14.06  14.12| 14.06   0.00|  0.00

                      Thanks once again!

                        Hi Irena
                        Looking at the POE device, it makes sense --> (1920×1080×1,5+2×1280×800×1)×8×2,2X ~ 90-95 Mbit/s.

                        What doesn't make sense is that you are only getting 14 fps at 1080p with a usb3 cable. Are you sure you didn't accidentally set the resolution to 4k on the Oak-D?

                        Thanks,
                        Jaka

                          Hi @jakaskerl

                          I'm a bit confused with the formula, this is the same as a link before? Therefor this make sanse, because as we seen my pc is able to fit no more than 100mbps network.

                          I made a significant improvement in the performance of the OAK-D camera by adjusting the pipeline settings. Previously, I had set the pipeline to 4K resolution, but I have now changed it to 1080P. When I ran the cam_test.py script again, I observed much better results, like you say. Here is the current output:

                          Connected cameras:
 -socket RGB   : IMX378 4056 x 3040 focus:auto  - COLOR
 -socket LEFT  : OV9282 1280 x  800 focus:fixed - MONO
 -socket RIGHT : OV9282 1280 x  800 focus:fixed - MONO
USB speed: SUPER
IR drivers: []
Cam:      rgb           left          right         camd     [host | capture timestamp]
FPS:  30.00| 30.00  30.00| 30.00  30.00| 30.00   0.00|  0.00

                          Thank you for your assistance, Your help has allowed me to gain a better understanding of how it's works.

                          Thanks a lot!

                            Hi Irena
                            Yep, it's the same formula just refactored so it looks cleaner.

                            (width1*height*n_pix_NV12 + n_mono_cameras*width2*height2*n_pix_mono) * bits_per_pix * FPS

                            Thanks,
                            Jaka