YOLO26-Depth

This repository provides an Axera NPU inference deployment example for the Ultralytics YOLO26-Depth depth-estimation model, including model conversion and AX8850N Python inference examples.

This version of YOLO26-Depth has been converted to run on the Axera 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

• AX8850N/AX8850
  • M4N-Dock(爱芯派Pro)
  • M.2 Accelerator card
• AX8910
• AX630C
  • 爱芯派2
  • Module-LLM
  • LLM630 Compute Kit
• AX615

Performance Statistics

AX8850N (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

AX8910

模型	分辨率	耗时(ms)	FPS	CMM(MiB)
yolo26n-depth	768x768	20.42	48.97	16.38
yolo26s-depth	768x768	29.24	34.20	25.27
yolo26m-depth	768x768	52.34	19.11	48.02
yolo26l-depth	768x768	66.61	15.01	45.65
yolo26x-depth	768x768	120.21	8.32	83.85

AX630C (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

AX615 (NPU2 mode)

模型	分辨率	耗时(ms)	FPS	CMM(MiB)
yolo26n-depth	768x768	82.17	12.17	14.21
yolo26s-depth	768x768	109.03	9.17	23.44
yolo26m-depth	768x768	170.99	5.85	47.75
yolo26l-depth	768x768	201.56	4.96	45.57

How to use

Download all files from this repository to the device

root@ax650:~/root/yolo26-depth# tree -L 2
.
├── asserts
│   ├── bus.jpg
│   ├── ssd_car_depth_heatmap_ax.png
│   ├── ssd_car_depth_heatmap_onnx.png
│   └── ssd_car.jpg
├── ax615
│   ├── 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
├── ax630c
│   ├── 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
├── ax637
│   ├── 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
├── ax8850n
│   ├── 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

6 directories, 39 files

Inference with AX650 Host, such as M4N-Dock(爱芯派Pro)

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

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

# axera npu infer
python3 infer_depth.py --model ./ax8850n/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

axera npu output: