yolov8-tflite-imx 1.0.7

YOLOv8 INT8 TFLite webcam inference for i.MX8MP (NPU / CPU)

yolov8-tflite-imx

YOLOv8 INT8 TFLite webcam inference for i.MX8MP (phyBOARD-Pollux).
Supports CPU and NPU acceleration via the VeriSilicon VX delegate.

---

Included models

Command	Model task	Default model path
detect	Object detection (COCO)	yolov8n_saved_model/yolov8n_full_integer_quant.tflite
pose	Pose estimation	yolov8n-pose_saved_model/yolov8n-pose_full_integer_quant.tflite
obb	Oriented bounding box	yolov8n-obb_saved_model/yolov8n-obb_full_integer_quant.tflite
segment	Instance segmentation	yolov8n-seg_saved_model/yolov8n-seg_full_integer_quant.tflite

---

Installation

On the i.MX8MP board (recommended)

# 1. Install tflite_runtime wheel for your Python / board combination
#    (download the correct .whl from https://github.com/google-coral/pycoral/releases
#     or build from source for your kernel)
pip install tflite_runtime-*.whl

# 2. Install this package
pip install yolov8-tflite-imx

From source (development)

git clone https://github.com/yourname/yolov8-tflite-imx.git
cd yolov8-tflite-imx
pip install -e .

---

Usage

Object detection

# CPU
detect --cam 0 --model path/to/yolov8n_full_integer_quant.tflite

# NPU (VX delegate)
detect --cam 0 --npu --model path/to/yolov8n_full_integer_quant.tflite

# Custom delegate path
detect --cam 0 --delegate /usr/lib/libvx_delegate.so --model path/to/model.tflite

Pose estimation

pose --cam 0                          # CPU
pose --cam 0 --npu                    # NPU

Oriented bounding box (OBB)

obb --cam 0                           # CPU
obb --cam 0 --npu                     # NPU

Instance segmentation

segment --cam 0                       # CPU
segment --cam 0 --npu                 # NPU

---

All flags (same for every command)

Flag	Default	Description
--model PATH	see table above	Path to .tflite model
--cam INT	0	Camera device ID
--conf FLOAT	0.25	Confidence threshold
--iou FLOAT	0.45	NMS IoU threshold
--npu	off	Enable NPU via /usr/lib/libvx_delegate.so
--delegate PATH	None	Custom delegate .so (overrides --npu)
--no-gst	off	Disable GStreamer, use V4L2 directly

Keys while running

Key	Action
q or ESC	Quit
s	Save current frame as capture_<task>.jpg

---

Requirements

• Python ≥ 3.8
• tflite_runtime (installed separately for your board/arch)
• opencv-python ≥ 4.5
• numpy ≥ 1.21
• NPU: /usr/lib/libvx_delegate.so present on the board

---

License

MIT