DeepBBS

Introduction

This repository contains python scripts for training and testing DeepBBS.

DeepBBS is a method for estimating the rigid transformation between two 3D point clouds. It is based on Best Buddies. The Best Buddies criterion is a strong indication for correct matches that, in turn, leads to accurate registration. Instead of finding best buddies in the input 3D space, a neural network is trained to find an embedding space, in which the Best Buddies Similarity measure is computed. Experiments show that DeepBBS is robust to occlusions, has a very large basin of attractions, and achieves state-of-the-art results on several datasets.

For technical details, please refer to:

DeepBBS: Deep Best Buddies for Point Cloud Registration (3DV 2021 oral paper).

Configuration

System requirments:

• python 3.7
• pytorch=1.5.1
• h5py
• scipy=1.5.0
• scikit-learn=0.23.2
• tqdm

or it can be installed with the environment.yml file:

conda env create -f environment.yml

Testing

In every test, for testing DeepBBS++ use the argument --DeepBBS_pp. For testing DeepBBS use --DeepBBS instead.

Unseen Point Clouds

Weights can be downloaded from here.

python main.py --n_subsampled_points=768 --DeepBBS_pp --model_path=./pretrained/unseen_point_clouds.t7 --eval

Unseen Categories

Weights can be downloaded from here.

python main.py --n_subsampled_points=768 --DeepBBS_pp --unseen=True --model_path=./pretrained/unseen_categories.t7 --eval

Gaussian Noise

Weights can be downloaded from here.

python main.py --n_subsampled_points 768 --DeepBBS_pp --gaussian_noise=True --model_path=./pretrained/gaussian_noise.t7 --eval

Different Samplings

Weights can be downloaded from here.

python main.py --DeepBBS_pp=True --different_pc=True --model_path=./pretrained/different_samplings.t7 --eval

Training

In every case, for training DeepBBS++ use the argument --DeepBBS_pp. For training DeepBBS use --DeepBBS instead.

Unseen Point Clouds

python main.py --n_subsampled_points=768 --exp_name=unseen_point_clouds

Unseen Categories

python main.py --n_subsampled_points=768 --unseen=True --exp_name=unseen_categories

Gaussian Noise

python main.py --n_subsampled_points 768 --gaussian_noise=True --exp_name=gaussian_noise

Different Samplings

python main.py --different_pc=True --exp_name=different_samplings