Change Prototype code to my code on prototype to run at 120 FPS

JosephSchramm

Hi Luxonis community,

I'm the founder of StrikeSure, an AI-powered ball-strike adjudication system for youth and amateur baseball. We've had a great initial Zoom call with your engineering team and they directed me here for follow-up technical questions.

We're running our current working prototype on an OAK-D S2 at 30fps and have achieved 88.67% accuracy across 300 throws. Our next major milestone is integrating our algorithm onto an OAK-D running at 120fps (we're calling it "Frankie" internally).

We're also working with OAK FFC OV9782 FF cameras and have a few specific questions:

1. **Stereo depth configuration:** Our optimal detection zone is approximately 3–4 feet from the camera (the strike zone in front of home plate). What stereo depth settings do you recommend for maximizing accuracy at that specific range?

2. **Minimal pipeline configuration:** We don't need ToF, thermal, HDR, YOLO, or most of the features in the standard cam_test.py script. What's the leanest, most reliable pipeline config you'd recommend for our use case — stereo depth at high fps feeding a custom CV algorithm?

3. **OAK FFC OV9782 FF compatibility:** Should we be building our pipeline specifically around the OV9782 FF camera specs, or can we largely adapt existing OAK-D pipeline code? Are there any known gotchas with this sensor for outdoor, high-speed applications?

4. **120fps stereo depth:** What are the tradeoffs or configuration requirements to achieve reliable stereo depth at 120fps vs. our current 30fps setup?

Any guidance from the community or Luxonis engineers would be hugely appreciated. Happy to share more technical detail about our use case if helpful.

Thank you!

OskarSonc

Hi JosephSchramm - few practical recommendations for this use case, love the name Frankie hehe :

For a 3-4 ft strike-zone window, I’d start with standard stereo plus subpixel, not Extended Disparity. Extended Disparity is mainly for pushing MinZ closer, while your target zone is already in a comfortable range for OAK-D-class stereo. In a controlled setup, it also helps to constrain the valid depth range to just the strike zone and keep temporal filtering off for a fast-moving ball.
For the leanest pipeline, keep it to stereo pair -> StereoDepth -> your custom CV. I would avoid the heavier demo-style configs and only add post-processing if you see a specific artifact. For 120 fps, the practical starting point on OAK-D-class hardware is 400P with a minimal stereo config, then add accuracy features back one at a time.
For OAK FFC OV9782 FF, you can mostly adapt existing OAK-D stereo pipeline patterns. The bigger factors are your final baseline, FOV, calibration quality, and exposure settings. For outdoor high-speed work, the main gotchas are calibration, keeping exposure short enough to control motion blur, and validating sunlight behavior on the exact lens/sensor combination you’ll ship.

Here are few link I think will help:

Stereo depth tuning guide: https://docs.luxonis.com/hardware/platform/depth/configuring-stereo-depth/
Depth accuracy reference: https://docs.luxonis.com/hardware/platform/depth/depth-accuracy/
FFC deployment guide: https://docs.luxonis.com/hardware/platform/deploy/ffc

Thanks,
Oskar