Skip to content

Latest commit

 

History

History
163 lines (123 loc) · 6.81 KB

README.md

File metadata and controls

163 lines (123 loc) · 6.81 KB
Raw

This is an unofficial project to improve and test PointMLP's performance for scene segmentation. Originally, PointMLP was developed for 3d object classification and segmentation on the ShapeNet dataset - outperforming heavyweight transformer models.

We - Friedrich Dang, Begüm Altunbas, Görkem Güzeler and Han Keçeli tested several modifications of PointMLP on Stanford's 3D Indoor Scene Dataset (S3DIS). We improved its performance by 3-5% across various categories by exchanging the non-learnable max-pooling aggregation function with self-attention layers for the global feature extraction. Together with an MLP, feature density is kept high and a former upsampling procedure was made redundant.

Some test results on S3DIS for some of our modifications are summarized below.

Modification Accuracy Instance mIOU Class mIOU
None 0.67 0.65 0.52
Random Sampling 0.68 0.67 0.53
Poisson + FPS Sampling 0.68 0.64 0.51
GAM Non-Linear Normalization 0.61 0.59 0.47
Self-Attention global context 0.70 0.69 0.56
Multihead Self-Attention - color and global context 0.70 0.69 0.57

Install

conda create -n pointmlp python=3.7 -y
conda activate pointmlp
conda install pytorch==1.10.1 torchvision==0.11.2 cudatoolkit=10.2 -c pytorch -y
# if you are using Ampere GPUs (e.g., A100 and 30X0), please install compatible Pytorch and CUDA versions, like:
# pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
pip install cycler einops h5py pyyaml==5.4.1 scikit-learn==0.24.2 scipy tqdm matplotlib==3.4.2
pip install pointnet2_ops_lib/.

Useage

Scene segmentation S3DIS

  • Make data folder under part_segmentation/data

  • Download our subsample of S3DIS dataset from https://drive.google.com/drive/folders/1CW5qicxlE5tkpRiczTI84qDZxQojymD0?usp=share_link

  • Download the S3DIS dataset to that folder and save it as custom_s3 (arrange root path in main.py accordingly).

  • To perform evaluation please download model checkpoints from the same drive link and place it under /checkpoints folder (arrange paths accoringly).

  • Train & Test pointMLP

# train pointMLP
cd part_segmentation
python main.py --model pointMLP
# please add other paramemters as you wish.
# eval pointMLP
python main.py --model pointMLP --eval True
  • Visualize pointMLP on S3DIS dataset
# visualize pointMLP
cd part_segmentation
python visualize_test.py --log_dir pointmlp --test_area 5 --visual
# make sure that you have log/sem_seg/pointmlp/visual & log/sem_seg/pointmlp/logs folders created and correct checkpoint in checkpoints folder.

visualize_test.py saves .obj files inside log folder for 11 test instances with gt and predictions. To visualize these .obj files we have another script open3d_show_obj.py (arrange the .obj file path accordingly). Beware that Meshlab shows all points in white color thus it is recomended to use open3d_show_obj.py. Example .obj files are placed in the google drive link.

P.S: The data preprocessing (creating custom_s3/x.npy files), visualization and dataloader codes are based on https://github.com/yanx27/Pointnet_Pointnet2_pytorch and modified according to our project needs.

Part segmentation ShapeNet

  • Make data folder and download the dataset
cd part_segmentation
mkdir data
cd data
wget https://shapenet.cs.stanford.edu/media/shapenetcore_partanno_segmentation_benchmark_v0_normal.zip --no-check-certificate
unzip shapenetcore_partanno_segmentation_benchmark_v0_normal.zip
  • Train pointMLP
# train pointMLP
python main-original.py --model pointMLP
# please add other paramemters as you wish.

Classification ModelNet40

Train: The dataset will be automatically downloaded, run following command to train.

By default, it will create a folder named "checkpoints/{modelName}-{msg}-{randomseed}", which includes args.txt, best_checkpoint.pth, last_checkpoint.pth, log.txt, out.txt.

cd classification_ModelNet40
# train pointMLP
python main.py --model pointMLP
# train pointMLP-elite
python main.py --model pointMLPElite
# please add other paramemters as you wish.

To conduct voting testing, run

# please modify the msg accrodingly
python voting.py --model pointMLP --msg demo

Classification ScanObjectNN

The dataset will be automatically downloaded

  • Train pointMLP/pointMLPElite
cd classification_ScanObjectNN
# train pointMLP
python main.py --model pointMLP
# train pointMLP-elite
python main.py --model pointMLPElite
# please add other paramemters as you wish.

By default, it will create a fold named "checkpoints/{modelName}-{msg}-{randomseed}", which includes args.txt, best_checkpoint.pth, last_checkpoint.pth, log.txt, out.txt.

Acknowledgment

Our implementation is mainly based on the PointMLP codebase. We gratefully thank the authors for their wonderful works.

S3DIS

@ARTICLE{2017arXiv170201105A,
   author = {{Armeni}, I. and {Sax}, A. and {Zamir}, A.~R. and {Savarese}, S.
	},
    title = "{Joint 2D-3D-Semantic Data for Indoor Scene Understanding}",
  journal = {ArXiv e-prints},
archivePrefix = "arXiv",
   eprint = {1702.01105},
 primaryClass = "cs.CV",
 keywords = {Computer Science - Computer Vision and Pattern Recognition, Computer Science - Robotics},
     year = 2017,
    month = feb,
   adsurl = {http://adsabs.harvard.edu/abs/2017arXiv170201105A},
  adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

PointMLP

PWC PWC

github

open review | arXiv | Primary contact: Xu Ma

LICENSE

PointMLP is under the Apache-2.0 license.