This is a repository forked from Hugues THOMAS/KPConv-PyTorch. In this repository, KPConv is used as an encoder to learn the local features of the point clouds. NetVLAD is intergrated to learn the global descriptor for performing place recognition.
This implementation has been tested on Ubuntu 20.04 with docker evironment. Details are provided in docker folder.
This implementation has been tested on Oxford Robocar benchmark which is created by mikacuy/pointnetvlad. By training it on 3 sequences, the average recall is 69.55 %.
Training:
python train_Oxford.py
Evaluation:
python evaluation_place_recognition.py
Results:
