RTMPose-M

This version of RTMPose-M (256x192) has been converted to run on the Axera NPU using mixed w16/fp32 quantization. It is optimized for real-time human pose estimation with 17 COCO keypoints using the SimCC decoding approach.

Compatible with Pulsar2 version: 6.0.

Model Info

Item	Value
Architecture	RTMPose-M (CSPNeXt + SimCC Head)
Parameters	13.58M
Input	1x256x192x3 (NHWC, uint8, BGR)
Output	simcc_x (1,17,384), simcc_y (1,17,512)
Keypoints	17 (COCO format)
Source	OpenMMLab MMPose

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

Performance Statistics

AX650N

Model	Latency(ms) npu3
rtmpose_m	2.881

Conversion Pipeline

1. Export ONNX — Download official RTMPose-M from OpenMMLab and fix batch dim:

python export_onnx.py

2. Replace HardSigmoid — Replace HardSigmoid ops with Mul+Add+Clip for better NPU quantization:

python replace_hardsigmoid.py

3. Compile axmodel — Use Pulsar2 with the provided config.json to quantize and compile:

pulsar2 build --target_hardware AX650 --config config.json --input rtmpose_m_256x192_no_hs.onnx --output_dir AX650

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

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

Input image:

run

python3 ax_infer.py -m rtmpose_m_npu3.axmodel -i test.jpg

root@ax650:~/data# python3 ax_infer.py -m rtmpose_m_npu3.axmodel -i test.jpg
[INFO] Available providers:  ['AxEngineExecutionProvider']
[INFO] Using provider: AxEngineExecutionProvider
[INFO] Chip type: ChipType.MC50
[INFO] VNPU type: VNPUType.DISABLED
[INFO] Engine version: 2.10.1s
[INFO] Model type: 2 (triple core)
[INFO] Compiler version: 6.0 93b95f7f
Model input: name=input, shape=[1, 256, 192, 3], dtype=uint8
Forward: 3.38 ms (avg of 10 runs)
simcc_x: shape=(1, 17, 384), range=[-0.58, 0.88]
simcc_y: shape=(1, 17, 512), range=[-0.49, 0.88]
kpts above 0.3: 17/17
  kp00: ( 359.6,   83.3)  score=0.6773
  kp01: ( 370.0,   79.2)  score=0.6950
  kp02: ( 359.6,   77.1)  score=0.6878
  kp03: ( 384.6,   79.2)  score=0.7398
  kp04: ( 359.6,   79.2)  score=0.6385
  kp05: ( 403.3,  106.3)  score=0.7596
  kp06: ( 367.9,  116.7)  score=0.7683
  kp07: ( 432.5,  152.1)  score=0.4699
  kp08: ( 342.9,  158.3)  score=0.6831
  kp09: ( 445.0,  177.1)  score=0.3021
  kp10: ( 305.4,  179.2)  score=0.5798
  kp11: ( 432.5,  212.5)  score=0.7872
  kp12: ( 399.2,  218.7)  score=0.8110
  kp13: ( 432.5,  289.6)  score=0.7358
  kp14: ( 372.1,  279.2)  score=0.8252
  kp15: ( 470.0,  356.2)  score=0.6704
  kp16: ( 399.2,  345.8)  score=0.8183
Saved: ax_result.jpg

Output image: