Skip to main content

imx8mp-npu-pose — single-person pose estimation on the i.MX8M Plus NPU

View source on GitHub

Prerequisites: Avocado CLI (>=0.26.0), Docker, a ucm-imx8m-plus board with a USB webcam, and the eIQ ML stack in your feed (packagegroup-avocado-imx-ml). A display is optional — the browser preview works headless.

1. Fetch the model + calibration images​

./fetch-model.sh          # MoveNet SavedModel -> app/build/movenet, ~40 images -> app/build/rep

For real accuracy, drop frames with people (in your scene's lighting/poses) into app/build/rep/ and re-run the build.

2. Build (quantization happens here, in the SDK)​

avocado build

app-compile.sh runs in the SDK: it uv pip installs TensorFlow 2.16, runs quantize-model.py to produce a per-tensor INT8 movenet_int8.tflite, and app-install.sh copies it into the app extension. TensorFlow is not shipped to the target.

3. Provision​

avocado provision -r dev      # then flash per your board (uuu-emmc / sd)

4. View it​

  • Browser (headless or not): open http://<board-ip>:8080/ — live camera with the skeleton overlaid.
  • HDMI display: if a KMS display + weston are up, the skeleton renders there automatically too.
  • Journal: journalctl -fu imx8mp-npu-pose shows FPS + visible-keypoint count.

Confirm the NPU is doing the work:

lsmod | grep galcore 2>/dev/null; ls -l /dev/galcore     # NPU present
journalctl -u imx8mp-npu-pose | grep -i "vx delegate"    # error_during_* = 0

5. NPU vs CPU​

systemctl set-environment USE_NPU=1 && systemctl restart imx8mp-npu-pose   # NPU
systemctl set-environment USE_NPU=0 && systemctl restart imx8mp-npu-pose   # CPU

The INT8 MoveNet should run materially faster on the NPU. Equal FPS ⇒ the model fell back to CPU (check the journal for VX-delegate errors / per-op fallback).

Troubleshooting​

  • No skeleton, "0/17 keypoints": no person in frame, or SCORE_THRESH too high (systemctl set-environment SCORE_THRESH=0.2). Stand back so your whole body is in the (square-cropped) frame.
  • Wrong camera: auto-detect picks the first real capture node; override with CAMERA_DEVICE=/dev/videoN.
  • Skeleton offset/mirrored: coordinate mapping assumes the square crop — if you change PREVIEW_SIZE or the crop, re-check the mapping in on_draw.
  • Black screen on HDMI: that's the display path, not pose — see the imx8mp-npu-nnstreamer notes (needs a KMS display + weston). The web preview sidesteps it entirely.