This repository contains the SuperPoint model converted to run on the Axera NPU. SuperPoint is a self-supervised framework for training interest point detectors and descriptors, suitable for a large number of multiple-view geometry problems in computer vision.

This version has been quantized using w8a16 and is optimized for use with Pulsar2 (version 4.2).

Convert Tools Links

For developers interested in custom model conversion or optimization:

AXera Platform GitHub Repo: Comprehensive guides and sample code for AXera chips.
Pulsar2 Documentation: Official guide on converting ONNX models to .axmodel.

Supported Platforms

AX650
AX637

Chips	SuperPoint (640x480)
AX650	27.443 ms
AX637	96.118 ms

How to Use

Download the model files and the inference binary to your Axera-powered device:

bash

root@ax650:~/SuperPoint-Demo# tree
.
|-- ax650
|   `-- compiled.axmodel
|-- infer.py
|-- 1.ppm
|-- 2.ppm
`-- output.jpg

Inference

python3 infer.py --model ./ax650/compiled.axmodel --img1 1.ppm --img2 2.ppm --output output.jpg

AXERA-TECH/superpoint

SuperPoint

Convert Tools Links

Supported Platforms

How to Use

Inference

Output image

AXERA-TECH/
superpoint