LightGlue: Optimized for Qualcomm Devices

LightGlue is a deep neural network that matches sparse local features across image pairs. It replaces the traditional combination of nearest-neighbour matching and robust geometric fitting with a fast, lightweight attention-based model, performing both matching and a learned pruning of non-matchable points in a single forward pass. LightGlue consumes keypoints and descriptors already produced by a separate feature extractor (SuperPoint) and outputs point correspondences between the two images. This recipe pairs LightGlue's official checkpoint (trained against Magic-Leap's SuperPoint) with this repo's MIT-licensed SuperPoint port, so reported accuracy is directionally meaningful but not a direct reproduction of the paper's figures.

This is based on the implementation of LightGlue found here. This repository contains pre-exported model files optimized for Qualcomm® devices. You can use the Qualcomm® AI Hub Models library to export with custom configurations. More details on model performance across various devices, can be found here.

Qualcomm AI Hub Models uses Qualcomm AI Hub Workbench to compile, profile, and evaluate this model. Sign up to run these models on a hosted Qualcomm® device.

Getting Started

There are two ways to deploy this model on your device:

Option 1: Download Pre-Exported Models

Below are pre-exported model assets ready for deployment.

Runtime Precision Chipset SDK Versions Download
ONNX float Universal QAIRT 2.51, ONNX Runtime 1.30.0 Download
TFLITE float Universal QAIRT 2.51 Download

For more device-specific assets and performance metrics, visit LightGlue on Qualcomm® AI Hub.

Option 2: Export with Custom Configurations

Use the Qualcomm® AI Hub Models Python library to compile and export the model with your own:

  • Custom weights (e.g., fine-tuned checkpoints)
  • Custom input shapes
  • Target device and runtime configurations

This option is ideal if you need to customize the model beyond the default configuration provided here.

See our repository for LightGlue on GitHub for usage instructions.

Model Details

Model Type: Model_use_case.pose_estimation

Model Stats:

  • Descriptor dimension: 256
  • Input resolution (pixel coords): 480x640
  • Max input keypoints per image: 500
  • Model checkpoint: superpoint_lightglue_v0-1_arxiv.pth
  • Number of transformer layers: 9
  • Number of parameters: 11.3M
  • Model size (float): 43.7 MB

Performance Summary

Model Runtime Precision Chipset Inference Time (ms) Peak Memory Range (MB) Primary Compute Unit
lightglue ONNX float Snapdragon® 8 Elite Gen 5 For Galaxy Mobile 48.077 ms 23 - 402 MB NPU
lightglue ONNX float Snapdragon® 8 Elite For Galaxy Mobile 46.805 ms 22 - 385 MB NPU
lightglue ONNX float Snapdragon® X2 Elite 59.137 ms 32 - 32 MB NPU
lightglue ONNX float Snapdragon® X Elite 118.295 ms 32 - 32 MB NPU
lightglue ONNX float Qualcomm® Dragonwing™ IQ-X7181 118.295 ms 32 - 32 MB NPU
lightglue ONNX float Qualcomm® Dragonwing™ Q-8750 46.805 ms 22 - 385 MB NPU
lightglue ONNX float Snapdragon® 8 Elite Extreme Gen 6 Mobile 36.05 ms 23 - 356 MB NPU
lightglue ONNX float Snapdragon® 8 Elite Gen 6 Mobile 59.968 ms 23 - 408 MB NPU
lightglue TFLITE float Snapdragon® 8 Elite Gen 5 For Galaxy Mobile 49.274 ms 14 - 834 MB NPU
lightglue TFLITE float Snapdragon® 8 Elite For Galaxy Mobile 55.718 ms 14 - 805 MB NPU
lightglue TFLITE float Snapdragon® 8 Gen 1 Mobile 118.9 ms 15 - 1052 MB NPU
lightglue TFLITE float Qualcomm® Dragonwing™ IQ-8275 103.289 ms 14 - 80 MB NPU
lightglue TFLITE float Qualcomm® Dragonwing™ QCS8550 (Proxy) 103.447 ms 14 - 19 MB NPU
lightglue TFLITE float Qualcomm® SA8775P 116.398 ms 15 - 800 MB NPU
lightglue TFLITE float Qualcomm® SA8650P 116.398 ms 15 - 800 MB NPU
lightglue TFLITE float Qualcomm® SA8255P 116.398 ms 15 - 800 MB NPU
lightglue TFLITE float Qualcomm® QCS8450 118.9 ms 15 - 1052 MB NPU
lightglue TFLITE float Qualcomm® Dragonwing™ IQ-9075 110.159 ms 12 - 78 MB NPU
lightglue TFLITE float Qualcomm® Dragonwing™ Q-8750 55.718 ms 14 - 805 MB NPU
lightglue TFLITE float Qualcomm® SA8295P 124.412 ms 15 - 800 MB NPU
lightglue TFLITE float Snapdragon® 8 Elite Extreme Gen 6 Mobile 38.452 ms 14 - 700 MB NPU
lightglue TFLITE float Snapdragon® 8 Elite Gen 6 Mobile 58.72 ms 13 - 688 MB NPU
superpoint ONNX float Snapdragon® 8 Elite Gen 5 For Galaxy Mobile 6.836 ms 28 - 199 MB NPU
superpoint ONNX float Snapdragon® 8 Elite For Galaxy Mobile 9.353 ms 28 - 195 MB NPU
superpoint ONNX float Snapdragon® X2 Elite 7.468 ms 40 - 40 MB NPU
superpoint ONNX float Snapdragon® X Elite 16.773 ms 40 - 40 MB NPU
superpoint ONNX float Qualcomm® Dragonwing™ IQ-X7181 16.773 ms 40 - 40 MB NPU
superpoint ONNX float Qualcomm® Dragonwing™ Q-8750 9.353 ms 28 - 195 MB NPU
superpoint ONNX float Snapdragon® 8 Elite Extreme Gen 6 Mobile 5.649 ms 34 - 202 MB NPU
superpoint ONNX float Snapdragon® 8 Elite Gen 6 Mobile 6.698 ms 33 - 198 MB NPU
superpoint TFLITE float Snapdragon® 8 Elite Gen 5 For Galaxy Mobile 6.792 ms 10 - 175 MB NPU
superpoint TFLITE float Snapdragon® 8 Elite For Galaxy Mobile 9.318 ms 11 - 177 MB NPU
superpoint TFLITE float Snapdragon® 8 Gen 1 Mobile 28.152 ms 10 - 246 MB NPU
superpoint TFLITE float Qualcomm® Dragonwing™ IQ-8275 28.47 ms 0 - 25 MB NPU
superpoint TFLITE float Qualcomm® Dragonwing™ QCS8550 (Proxy) 16.709 ms 12 - 25 MB NPU
superpoint TFLITE float Qualcomm® SA8775P 23.865 ms 12 - 173 MB NPU
superpoint TFLITE float Qualcomm® SA8650P 23.865 ms 12 - 173 MB NPU
superpoint TFLITE float Qualcomm® SA8255P 23.865 ms 12 - 173 MB NPU
superpoint TFLITE float Qualcomm® QCS8450 28.152 ms 10 - 246 MB NPU
superpoint TFLITE float Qualcomm® Dragonwing™ IQ-9075 24.507 ms 0 - 24 MB NPU
superpoint TFLITE float Qualcomm® Dragonwing™ Q-8750 9.318 ms 11 - 177 MB NPU
superpoint TFLITE float Qualcomm® SA8295P 31.101 ms 12 - 189 MB NPU
superpoint TFLITE float Snapdragon® 8 Elite Extreme Gen 6 Mobile 5.456 ms 17 - 182 MB NPU
superpoint TFLITE float Snapdragon® 8 Elite Gen 6 Mobile 6.697 ms 10 - 174 MB NPU

License

  • The license for the original implementation of LightGlue can be found here.

References

Community

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Paper for qualcomm/LightGlue