YOLO26-Pose

This version of YOLOv26-Pose has been converted to run on the Axera NPU using w8a16 quantization.

Compatible with Pulsar2 version: 4.2.

Convert tools links:

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

• The repo of AXera Platform, where you can get the detailed guide.
• Pulsar2 Link, How to Convert ONNX to axmodel

Support Platform

• AX650N/AX8850
  • M4N-Dock(爱芯派Pro)
  • M.2 Accelerator card
• AX630C
  • 爱芯派2
  • Module-LLM
  • LLM630 Compute Kit
• AX615
• AX637

Performance Statistics

AX650N(NPU1)

Model	FPS	CMM(MB)	Latency(ms)
yolo26n-pose	240.21	5.71	4.163
yolo26s-pose	99.30	17.90	10.070
yolo26m-pose	36.96	37.82	27.053
yolo26l-pose	28.88	39.25	34.620
yolo26x-pose	13.50	79.61	74.072

AX650N(NPU3)

Model	FPS	CMM(MB)	Latency(ms)
yolo26n-pose	655.74	3.80	1.525
yolo26s-pose	283.45	11.27	3.528
yolo26m-pose	107.57	28.98	9.296
yolo26l-pose	83.59	35.22	11.963
yolo26x-pose	39.80	72.24	25.128

AX630C

Model	Latency(ms) npu1	Latency(ms) npu2
yolo26n-pose	11.003	7.203
yolo26s-pose	24.157	17.120

AX615

Model	Latency(ms) npu1	Latency(ms) npu2
yolo26n-pose	19.542	11.446
yolo26s-pose	57.481	28.553

AX637

Model	Latency(ms)
yolo26n-pose	4.595
yolo26s-pose	11.848

How to use

Download all files from this repository to the device

python env requirement

pyaxengine

https://github.com/AXERA-TECH/pyaxengine

wget https://github.com/AXERA-TECH/pyaxengine/releases/download/0.1.3.rc2/axengine-0.1.3-py3-none-any.whl
pip install axengine-0.1.3-py3-none-any.whl

others

Maybe None.

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

Input image:

run

python ax_infer_pose.py --model-path yolo26l-pose_npu3.axmodel --test-img bus.jpg

(ax_env) root@ax650:~/ax650# python ax_infer.py --model-path yolo26l-pose_npu3.axmodel --test-img bus.jpg
[INFO] Available providers:  ['AxEngineExecutionProvider']
[INFO] Using provider: AxEngineExecutionProvider
[INFO] Chip type: ChipType.MC50
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Engine version: 2.12.0s
[INFO] Model type: 2 (triple core)
[INFO] Compiler version: 5.1-patch1-dirty 9164b433-dirty
[YOLO26-Pose] [16:47:23.013] [DEBUG] Load model time = 838.91 ms
[YOLO26-Pose] [16:47:23.070] [DEBUG] Pre-process time = 16.87 ms
[YOLO26-Pose] [16:47:23.109] [DEBUG] Forward time = 39.14 ms
[YOLO26-Pose] [16:47:23.113] [DEBUG] Post-process time = 2.97 ms
[YOLO26-Pose] [16:47:23.117] [INFO] Draw Results (4 persons):
[YOLO26-Pose] [16:47:23.117] [INFO] (221, 406, 345, 859) -> person: 0.92
[YOLO26-Pose] [16:47:23.120] [INFO] (668, 391, 807, 880) -> person: 0.92
[YOLO26-Pose] [16:47:23.122] [INFO] (47, 397, 243, 901) -> person: 0.90
[YOLO26-Pose] [16:47:23.123] [INFO] (0, 439, 78, 949) -> person: 0.69
[YOLO26-Pose] [16:47:23.151] [INFO] Saved to result_yolo26_pose.jpg

Output image: