Minimal CVITEK CV1812H/Milk‑V Duo S demo - runs a single‑class YOLOv8 model on the on‑chip TPU and prints every detected deer. This is largely a patch of the SDK-provided demos, refactoring from the default 80 class input to 1.
Deer don't belong in my garden (they eat my tomatoes!) so the use case of this tool is using an onboard camera (GC2083), run a custom YOLOv8 model trained on pictures of whitetail + mule deer, and print detection results on a schedule.
git clone https://github.com/milkv-duo/duo-tdl-examples.git
cd duo-tdl-examples/
source envsetup.sh
Select your platform + arm64/riscv64 :
script_dir: /home/easto/duo-tdl-examples
Select Product:
1. Duo (CV1800B)
2. Duo256M (SG2002) or DuoS (SG2000)
Which would you like: 2
Select Arch:
1. ARM64
2. RISCV64
Which would you like: 2
CHIP: CV181X
ARCH: riscv64
~/duo-tdl-examples ~/duo-tdl-examples
~/duo-tdl-examples
Environment is ready.
then compile deer_img:
make
./deer_img <model.cvimodel> <image>
Image 1920×1080
Inference: 45 ms
[0] deer 0.92 (345,420)-(680,800)
Adjust confidence / NMS inside deer_img.c:
CVI_TDL_SetModelThreshold(h, 0.80f); // 80 minimum, 82-85 seems to work reliably
CVI_TDL_SetModelNmsThreshold(h, 0.45f); // 0.45 minimum, 0.5-0.6 work bestBuilt and tested with the CV181x (duo-buildroot-sdk-v2) examples repo
https://github.com/milkv-duo/duo-tdl-examples/tree/master