Multi-instance-Point-Cloud-Registration-by-Efficient-Correspondence-Clustering

This is the source code of our CVPR paper Arxiv

Install Environment:

conda env create -f multiregister.yaml

Test on synthetic dataset and real dataset (Scan2CAD, ModelNet40)

All the experimental code files are in ./synthetic&real

Weights

Download the weights and put multi_oneTomore_multi_1 and multi_real_box_test_main_cad directly into ./synthetic&real/snapshot

Datas

The Scan2CAD dataset may need to be downloaded on Scan2CAD, and put ./split.json in the dataset folder. If you choose ModelNet40 for synthetic experiments, then you may download the dataset in Data provided by Pointnet++ and put ./modelnet40_train.json,./modelnet40_test.json and modelnet40_classnum2label.json into the dataset folder.

Install pytorch extension:

cd ./synthetic&real 
cd cpp_wrappers 
sh build.sh 
cd ..

Run

There are two commands to conduct the two experiments respectively in run.sh, you can choose which to run.

Test on our rgbd data

All the experimental code files are in ./rgbd

Datas

Download the datas and put scenes and objects directly into ./rgbd

Run

sh run.sh to run and you can see the visualized results.