点云配准（RPM-Net: Robust Point Matching using Learned Features）源码

• 论文: RPM-Net: Robust Point Matching using Learned Features

 这是一篇2020年的论文

• 代码: GitHub - yewzijian/RPMNet: RPM-Net: Robust Point Matching using Learned Features (CVPR2020)

Prerequisites

See requirements.txt for required packages. Our source code was developed using Python 3.6 with PyTorch 1.2. However, we do not observe problems running on newer versions available as of time of writing.

Training

Run the relevant commands below. We use the processed ModelNet40 dataset from PointNet for this work, which will be automatically downloaded if necessary. Performance typically saturates by around 500-1000 epochs depending on setting.

mkdir rpmnet && cd rpmnet
git clone git@github.com:yewzijian/RPMNet.git
cd RPMNet/src
python train.py --noise_type crop

For clean and noisy data, we use a smaller batch size of 4 to allow it to train on a 11GB GPU (e.g. Nvidia GTX 1080Ti), but for the rest of the experiments we use a batch size of 8. So, for clean data, replace the last line with:

python train.py --noise_type clean --train_batch_size 4

, and for noisy data:

python train.py --noise_type jitter --train_batch_size 4

The tensorboard summaries and more importantly the checkpoints will be saved in [root]/logs/[datetime]/*. Note that you need a recent version of tensorboard if you wish to visualize the point clouds (optional).

Inference / Evaluation

This script performs inference on the trained model, and computes evaluation metrics.

Note: replace --noise_type accordingly if not running on partial data.

python eval.py --noise_type crop --resume [path-to-logs/ckpt/model-best.pth]

Alternatively, given transforms saved in a .npy file of shape (B, [n_iter], 3, 4), you can evaluate them using:

python eval.py --noise_type crop --transform_file [path-to-transform-file.npy]

Pretrained models

Our pretrained models can be downloaded from here. You should be able to obtain the results shown in the paper by using these checkpoints.

源码详解：未完待续....