YOLO26-Depth

This repository provides an WaveMatrix NPU inference deployment example for the Ultralytics YOLO26-Depth depth-estimation model, including model conversion and WM9955 Python inference examples.

This version of YOLO26-Depth has been converted to run on the WaveMatrix NPU using mixed U8/U16 quantization for improved performance.

Convert tools links:

For those who are interested in model conversion, you can try to export axmodel through

• Pulsar2 Link, How to Convert ONNX to axmodel

Support Platform

• WM9955
  • M.2 Accelerator card
• WM8845
  • Demo board

Performance Statistics

WM9955 (NPU3 mode)

模型	分辨率	耗时(ms)	FPS	CMM(MiB)
yolo26n-depth	768x768	5.83	171.53	7.97
yolo26s-depth	768x768	8.33	120.05	14.90
yolo26m-depth	768x768	16.37	61.09	36.36
yolo26l-depth	768x768	20.74	48.22	41.12
yolo26x-depth	768x768	39.19	25.52	93.37

WM8845 (NPU2 mode)

模型	分辨率	耗时(ms)	FPS	CMM(MiB)
yolo26n-depth	768x768	27.67	36.14	15.65
yolo26s-depth	768x768	41.68	23.99	22.75
yolo26m-depth	768x768	74.02	13.51	47.10
yolo26l-depth	768x768	94.04	10.63	44.77
yolo26x-depth	768x768	169.09	5.91	84.50

How to use

Download all files from this repository to the device

root@wm9955~/root/yolo26-depth# tree -L 2
.
├── asserts
│   ├── bus.jpg
│   ├── ssd_car_depth_heatmap_ax.png
│   ├── ssd_car_depth_heatmap_onnx.png
│   └── ssd_car.jpg
├── wm8845
│   ├── config.json
│   ├── yolo26l-depth_w8a8_mix.axmodel
│   ├── yolo26m-depth_w8a8_mix.axmodel
│   ├── yolo26n-depth_w8a8_mix.axmodel
│   ├── yolo26s-depth_w8a8_mix.axmodel
│   └── yolo26x-depth_w8a8_mix.axmodel
├── wm9955
│   ├── config.json
│   ├── yolo26l-depth_w8a8_mix.axmodel
│   ├── yolo26m-depth_w8a8_mix.axmodel
│   ├── yolo26n-depth_w8a8_mix.axmodel
│   ├── yolo26s-depth_w8a8_mix.axmodel
│   └── yolo26x-depth_w8a8_mix.axmodel
├── axengine-0.1.3-py3-none-any.whl
├── calib_data.zip
├── fold_depth_log.py
├── infer_depth.py
├── onnx
│   ├── yolo26l-depth.onnx
│   ├── yolo26m-depth.onnx
│   ├── yolo26n-depth.onnx
│   ├── yolo26s-depth.onnx
│   └── yolo26x-depth.onnx
├── README.md
└── requirements.txt

Inference with WM9955 Host, such as demo board

cd /root/yolo26-depth
pip3 install -r requirements.txt

pip3 install axengine-0.1.3-py3-none-any.whl

# wavematrix npu infer
python3 infer_depth.py --model ./wm9955/yolo26n-depth_w8a8_mix.axmodel --source ./asserts/ssd_car.jpg

# onnxtuntime cpu infer
python3 infer_depth.py --model ./onnx/yolo26n-depth.onnx --source ./asserts/ssd_car.jpg

wavematrix npu output: