Installation

Preparation

Requirements

• Ubuntu 16.04
• Anaconda with python=3.6
• pytorch>=1.3 (I used 1.4 since no 1.3 currently(not working on pytorch 1.5))
• torchvision with pillow<7 (intall this manually ahead of time)
• cuda=10.1
• others: pip install termcolor opencv-python tensorboard h5py easydict
• note

Datasets

Shape Classification on ModelNet40

You can download ModelNet40 for here (1.6 GB). Unzip and move (or link) it to data/ModelNet40/modelnet40_normal_resampled.

Part Segmentation on PartNet

You can download PartNet dataset from the ShapeNet official webpage (8.0 GB). Unzip and move (or link) it to data/PartNet/sem_seg_h5.

Part Segmentation on ShapeNetPart

You can download ShapeNetPart dataset from here (635M). Unzip and move (or link) it to data/ShapeNetPart/shapenetcore_partanno_segmentation_benchmark_v0.

Scene Segmentation on S3DIS

You can download the S3DIS dataset from here (4.8 GB). You only need to download the file named Stanford3dDataset_v1.2.zip, unzip and move (or link) it to data/S3DIS/Stanford3dDataset_v1.2.

The file structure should look like:

<pt-code-root>
├── cfgs
│   ├── modelnet
│   ├── partnet
│   └── s3dis
├── data
│   ├── ModelNet40
│   │   └── modelnet40_normal_resampled
│   │       ├── modelnet10_shape_names.txt
│   │       ├── modelnet10_test.txt
│   │       ├── modelnet10_train.txt
│   │       ├── modelnet40_shape_names.txt
│   │       ├── modelnet40_test.txt
│   │       ├── modelnet40_train.txt
│   │       ├── airplane
│   │       ├── bathtub
│   │       └── ...
│   ├── PartNet
│   │   └── sem_seg_h5
│   │       ├── Bag-1
│   │       ├── Bed-1
│   │       ├── Bed-2
│   │       ├── Bed-3
│   │       ├── Bottle-1
│   │       ├── Bottle-3
│   │       └── ...
│   ├── ShapeNetPart
│   │   └── shapenetcore_partanno_segmentation_benchmark_v0
│   │       ├── README.txt
│   │       ├── synsetoffset2category.txt
│   │       ├── train_test_split
│   │       ├── 02691156
│   │       ├── 02773838
│   │       ├── 02954340
│   │       ├── 02958343
│   │       ├── 03001627
│   │       ├── 03261776
│   │       └── ...
│   └── S3DIS
│       └── Stanford3dDataset_v1.2
│           ├── Area_1
│           ├── Area_2
│           ├── Area_3
│           ├── Area_4
│           ├── Area_5
│           └── Area_6
├── init.sh
├── datasets
├── function
├── models
├── ops
└── utils

Compile custom operators and pre-processing data

sh init.sh

Usage

Training

ModelNet

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> \
    function/train_modelnet_dist.py --cfg <config file> [--log_dir <log directory>]

• <port_num> is the port number used for distributed training, you can choose like 12347.
• <config file> is the yaml file that determines most experiment settings. Most config file are in the cfgs directory.
• <log directory> is the directory that the log file, checkpoints will be saved, default is log.

PartNet

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> \
    function/train_partnet_dist.py --cfg <config file> [--log_dir <log directory>]

ShapeNetPart

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> \
    function/train_shapenetpart_dist.py --cfg <config file> [--log_dir <log directory>]

S3DIS

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node <num_of_gpus_to_use> \
    function/train_s3dis_dist.py --cfg <config file> [--log_dir <log directory>]

Evaluating

For evaluation, we recommend using 1 gpu for more precise result.

ModelNet40

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node 1 \
    function/evaluate_modelnet_dist.py --cfg <config file> --load_path <checkpoint> [--log_dir <log directory>]

• <port_num> is the port number used for distributed evaluation, you can choose like 12347.
• <config file> is the yaml file that determines most experiment settings. Most config file are in the cfgs directory.
• <checkpoint> is the model checkpoint used for evaluating.
• <log directory> is the directory that the log file, checkpoints will be saved, default is log_eval.

PartNet

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node 1 \
    function/evaluate_partnet_dist.py --cfg <config file> --load_path <checkpoint> [--log_dir <log directory>]

ShapeNetPart

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node 1 \
    function/evaluate_shapenetpart_dist.py --cfg <config file> --load_path <checkpoint> [--log_dir <log directory>]

S3DIS

python -m torch.distributed.launch --master_port <port_num> --nproc_per_node 1 \
    function/evaluate_s3dis_dist.py --cfg <config file> --load_path <checkpoint> [--log_dir <log directory>]

Models

ModelNet40

Method	Acc	Model
Point-wise MLP	93.0	Google / Baidu(fj13)
Pseudo Grid	93.1	Google / Baidu(gmh5)
Adapt Weights	92.9	Google / Baidu(bbus)
PosPool	93.0	Google / Baidu(wuuv)
PosPool*	93.3	Google / Baidu(qcc6)

S3DIS

Method	mIoU	Model
Point-wise MLP	66.3	Google / Baidu(53as)
Pseudo Grid	65.0	Google / Baidu(8skn)
Adapt Weights	64.5	Google / Baidu(b7zv)
PosPool	65.5	Google / Baidu(z752)
PosPool*	65.5	Google / Baidu(r96f)

Data iteration indices: Google / Baidu(m5bp)

PartNet

Method	mIoU (val)	mIoU (test)	Model
Point-wise MLP	49.1	82.5	Google / Baidu(wxff)
Pseudo Grid	50.6	53.3	Google / Baidu(n6b7)
Adapt Weights	50.5	52.9	Google / Baidu(pc22)
PosPool	50.5	53.6	Google / Baidu(3qv5)
PosPool*	51.1	53.7	Google / Baidu(czyq)

ShapeNetPart

Method	mIoU	msIoU	Acc	Model
Point-wise MLP	85.7	84.1	94.5	Google / Baidu(mi2m)
Pseudo Grid	86.0	84.3	94.6	Google / Baidu(wde6)
Adapt Weights	85.9	84.5	94.6	Google / Baidu(dy1k)
PosPool	85.9	84.6	94.6	Google / Baidu(r2tr)
PosPool*	86.2	84.8	94.8	Google / Baidu(27ie)