results

This repository contains the code implemented for my dissertation at University of Edinburgh. The dissertation title is: On the Applicability of feature matching on embedded devices.

I will provide a description of the files below:

• profile.xlsx - SuperGlue only profiling on Google Cloud

• profile_agx_xavier.xlsx- SuperGlue only profiling on AGX Xavier

• profile_xavier_nx.xlsx - SuperGlue only profiling on Xavier NX

• profile_superpoint_gc.xlsx - SuperGlue and Superpoint profling on Google Cloud

• profile_superpoint_xaviernx.xlsx - SuperGlue and Superpoint profiling on Xavier NX

• profile_superpoint_agx_xavier.xlsx - SuperGlue and Superpoint profiling on AGX Xavier

• profile_compressed_xaviernx.xlsx - SuperGlue and Superpoint compressed model profiling on Xavier NX

• test.py - code used to profile SuperGlue only model on the devices

• match_pairs.py - code used to profile SuperGlue and Superpoint on different devices

• match_pars2.py - code used to profile SuperGlue and Superpoint compressed model on Xavier NX

• superpoint.py - superpoint code used in compressing the model

• superglue.py - superglue code used in compressing the model

• matching.py - class that coordinated everything in the pipeline and was used in compressing the model

Further instructions on configuring the devices:

• Get COCO Dataset Link: wget http://images.cocodataset.org/zips/train2014.zip

• Official SuperPoint + Superglue repo link: https://github.com/magicleap/SuperGluePretrainedNetwork . This was used for profiling SuperGlue and Superpoint combined model.

• SuperGlue repo link: https://github.com/skylook/SuperGlue . This was used to profile SuperGlue only model.

• Install Torch on Nvidia-Jetson: https://forums.developer.nvidia.com/t/pytorch-for-jetson-version-1-10-now-available/72048

• Configure TensorRT: https://docs.donkeycar.com/guide/robot_sbc/tensorrt_jetson_nano/