Official PyTorch implementation of PointNeXt, for the following paper:
PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies
by Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, Bernard Ghanem
TL;DR: In PointNeXt, we propose improved training and model scaling strategies to boost PointNet++ to the state-of-the-art level. PointNet++ with the proposed model scaling is named as PointNeXt, the next version of PointNets.
In the PointNeXt project, we propose a new and flexible codebase for point-based methods, namely OpenPoints. The biggest difference between OpenPoints and other libraries is that we focus more on reproducibility and fair benchmarking.
-
Extensibility: support many representative networks for point cloud understanding, such as PointNet, DGCNN, DeepGCN, PointNet++, ASSANet, PointMLP, and our PointNeXt. More networks can be built easily based on our framework since OpenPoints support a wide range of basic operations including graph convolutions, self-attention, farthest point sampling, ball query, e.t.c.
-
Reproducibility: all implemented models are trained on various tasks by at least three times. Pretrained models and logs are available.
-
Fair Benchmarking: in PointNeXt, we find a large part of performance gain is due to the training strategies. In this library, all models are trained with the improved training strategies and all achieve much higher accuracy than the original reported value.
-
Ease of Use: Build model, optimizer, scheduler, loss function, and data loader easily from cfg. Train and validate different models on various tasks by simply changing the
cfg\*\*.yaml
file.model = build_model_from_cfg(cfg.model) criterion = build_criterion_from_cfg(cfg.criterion)
Here is an example of
pointnet.yaml
(model configuration for PointNet model):model: NAME: BaseCls encoder_args: NAME: PointNetEncoder in_channels: 4 cls_args: NAME: ClsHead num_classes: 15 in_channels: 1024 mlps: [512,256] norm_args: norm: 'bn1d'
-
Online logging: Support wandb for checking your results anytime anywhere.
Install by
source install.sh
Note:
- the
install.sh
requires CUDA 11.1; if another version of CUDA is used,install.sh
has to be modified accordingly; - you might need read the
install.rst
for a step-by-step installation if the bash file (install.sh
) does not work for you by any chance; - for all experiments, we use wandb for online logging by default. Run
wandb --login
only at the first time in a new machine, or setwandn.use_wandb=False
if you do not want to use wandb. Read the official wandb documentation if needed.
Check README.md
file under cfgs
directory for detailed training and evaluation on each benchmark.
For example,
- Train and validate on ScanObjectNN for 3D object classification, check
cfgs/scanobjectnn/README.md
- Train and validate on S3DIS for 3D segmentation, check
cfgs/s3dis/README.md
Note:
- We use yaml to support training and validation using different models on different datasets. Just use
.yaml
file accordingly. For example, train on ScanObjectNN using PointNeXt:CUDA_VISIBLE_DEVICES=1 bash main_classification.sh cfgs/scanobjectnn/pointnext-s.yaml
, train on S3DIS using ASSANet-L:CUDA_VISIBLE_DEVICES=1 bash main_segmentation.sh cfgs/s3dis/assanet-l.yaml
.
We provide the training logs & pretrained models in column our released
trained with the improved training strategies proposed by our PointNeXt through anonymous Google shared links.
TP: Throughput (instance per second) measured using an NVIDIA Tesla V100 32GB GPU and a 32 core Intel Xeon @ 2.80GHz CPU.
Throughput is measured with 128 x 1024 points.
name | OA/mAcc (Original) | OA/mAcc (our released) | #params | FLOPs | Throughput (ins./sec.) |
---|---|---|---|---|---|
PointNet | 68.2 / 63.4 | 75.2 / 71.4 | 3.5M | 1.0G | 4212 |
DGCNN | 78.1 / 73.6 | 86.1 / 84.3 | 1.8M | 4.8G | 402 |
PointMLP | 85.4±1.3 / 83.9±1.5 | 87.7 / 86.4 | 13.2M | 31.4G | 191 |
PointNet++ | 77.9 / 75.4 | 86.2 / 84.4 | 1.5M | 1.7G | 1872 |
PointNeXt-S | 87.7±0.4 / 85.8±0.6 | 88.20 / 86.84 | 1.4M | 1.64G | 2040 |
Throughput (TP) is measured with 16 x 15000 points.
name | mIoU/OA/mAcc (Original) | mIoU/OA/mAcc (our released) | #params | FLOPs | TP |
---|---|---|---|---|---|
PointNet++ | 54.5 / 81.0 / 67.1 | 68.1 / 87.6 / 78.4 | 1.0M | 7.2G | 186 |
PointNeXt-S | 68.0 / 87.4 / 77.3 | 68.0 / 87.4 / 77.3 | 0.8M | 3.6G | 227 |
PointNeXt-B | 71.5 / 88.8 / 80.2 | 71.5 / 88.8 / 80.2 | 3.8M | 8.8G | 158 |
PointNeXt-L | 73.9 / 89.8 / 82.2 | 73.9 / 89.8 / 82.2 | 7.1M | 15.2G | 115 |
PointNeXt-XL | 74.9 / 90.3 / 83.0 | 74.9 / 90.3 / 83.0 | 41.6M | 84.8G | 46 |
Throughput (TP) is measured with 16 x 15000 points.
name | mIoU/OA/mAcc (Original) | mIoU/OA/mAcc (our released) | #params | FLOPs | TP |
---|---|---|---|---|---|
PointNet++ | 53.5 / 83.0 / - | 63.6 / 88.3 / 70.2 | 1.0M | 7.2G | 186 |
ASSANet | 63.0 / - /- | 65.8 / 88.9 / 72.2 | 2.4M | 2.5G | 228 |
ASSANet-L | 66.8 / - / - | 68.0 / 89.7/ 74.3 | 115.6M | 36.2G | 81 |
PointNeXt-S | 63.4±0.8 / 87.9±0.3 / 70.0±0.7 | 64.2 / 88.2 / 70.7 | 0.8M | 3.6G | 227 |
PointNeXt-B | 67.3±0.2 / 89.4±0.1 / 73.7±0.6 | 67.5 / 89.4 / 73.9 | 3.8M | 8.8G | 158 |
PointNeXt-L | 69.0±0.5 / 90.0±0.1 / 75.3±0.8 | 69.3 / 90.1 / 75.7 | 7.1M | 15.2G | 115 |
PointNeXt-XL | 70.5±0.3 / 90.6±0.2 / 76.8±0.7 | 71.1 / 91.0 / 77.2 | 41.6M | 84.8G | 46 |
The code and models of ShapeNetPart will come soon.
name | OA/mAcc (Original) | OA/mAcc (our released) | #params | FLOPs | Throughput (ins./sec.) |
---|---|---|---|---|---|
PointNet++ | 91.9 / - | 93.0 / 90.7 | 1.5M | 1.7G | 1872 |
PointNeXt-S (C=64) | 93.7±0.3 / 90.9±0.5 | 94.0 / 91.1 | 4.5M | 6.5G | 2033 |
More exampless are available in the paper.
This library is inspired by PyTorch-image-models and mmcv.
@Article{qian2022pointnext,
author = {Qian, Guocheng and Li, Yuchen and Peng, Houwen and Mai, Jinjie and Hammoud, Hasan and Elhoseiny, Mohamed and Ghanem, Bernard},
journal = {arXiv:2206.04670},
title = {PointNeXt: Revisiting PointNet++ with Improved Training and Scaling Strategies},
year = {2022},
}