CUE-Feature

Environment (Docker)

• Ubuntu 18.04
• PyTorch 1.12 + CUDA 11.3
• MinkowskiEngine 0.5.4

To set up the environment, modify TORCH_CUDA_ARCH_LIST in docker/Dockfile to match your GPU, and then run the following commands:

$local: docker build -t cue:1.0 docker
$local: docker run --gpus all --rm -itd --name cue -v /local_dir:/container_dir --shm-size 16G --ipc=host cue:1.0
$container: conda init
$(base)container: cd cue_feature/
$(base)container: ./install_env.sh

Dataset

To download the required datasets, run the following scripts:

./scripts/download_3dmatch.sh dbs/
./scripts/download_3dmatch_testbench.sh dbs/

Train CUE/CUE+

• First, to train FCGF , run the following command:

  python main.py  train=3dmatch_pair
• Then, to train CUE, populate weights in conf/train/3dmatch_pair_btl with the saved checkpoint path, for example: weights: logs/HCL_0419_161400/best_val_checkpoint.pth, and run the following command:

  python main.py train=3dmatch_pair_btl
• Alternatively, to train CUE+, populate weights in conf/train/3dmatch_pair_mbtl with the saved checkpoint path and run the following command:

  python main.py train=3dmatch_pair_mbtl

Evaluate CUE/CUE+

• Eval FCGF:

  python eval/eval_3dmatch.py  --model=[saved_checkpoint.pth] --extract_features=1    --evaluate_FMR=1
• Eval CUE:

  python eval/eval_3dmatch.py  --model=[saved_checkpoint.pth] --evaluate_ECE=1
• Eval CUE+:

  python eval/eval_3dmatch.py  --model=[saved_checkpoint.pth] --evaluate_ECE=1

Plot ECE of CUE/CUE+

• To plot the ECE of CUE/CUE+, run the following commands:

  python eval/ece_rg.py
  # populate the ece_results.pickle path in eval/plot.ece.py and then run
  python eval/plot_ece.py
  

Pretained models

Pretained models available at Dropbox.