Skip to content

Indoor Point cloud Segmentation on S3DIS

The models are trained on the subsampled point clouds (voxel size = 0.04). The model achieving the best performance on validation is selected to test on the original point clouds (not downsampled).

Dataset

Please cite the S3DIS paper [1] if you are going to use our presampled datasets. The presampling is just to collect all point clouds, area by area and room by room, following PointGroup.

mkdir -p data/S3DIS/
cd data/S3DIS
gdown https://drive.google.com/uc?id=1MX3ZCnwqyRztG1vFRiHkKTz68ZJeHS4Y
tar -xvf s3disfull.tar

Organize the dataset as follows:

data
 |--- S3DIS
        |--- s3disfull
                |--- raw
                      |--- Area_6_pantry_1.npy
                      |--- ...
                |--- processed
                      |--- s3dis_val_area5_0.040.pkl

Train

For example, train PointNext-XL 

CUDA_VISIBLE_DEVICES=0 python examples/segmentation/main.py --cfg cfgs/s3dis/pointnext-xl.yaml
* change the cfg file to use any other model, e.g. cfgs/s3dis/pointnet++.yaml for training PointNet++
* run the command at the root directory

Test on Area 5

Note testing is a must step since evaluation in training is performed only on subsampled point clouds not original point clouds. 

CUDA_VISIBLE_DEVICES=0 bash script/main_segmentation.sh cfgs/s3dis/pointnext-xl.yaml wandb.use_wandb=False mode=test --pretrained_path pretrained/s3dis/pointnext-xl/pointnext-xl-area5/checkpoint/pointnext-xl_ckpt_best.pth
* add visualize=True to save segmentation results as .obj files

Test on All Areas

CUDA_VISIBLE_DEVICES=0 python examples/segmentation/test_s3dis_6fold.py --cfg cfgs/s3dis/pointnext-xl.yaml mode=test --pretrained_path pretrained/s3dis/pointnext-xl

Profile Parameters, FLOPs, and Throughput

CUDA_VISIBLE_DEVICES=0 python examples/profile.py --cfg cfgs/s3dis/pointnext-xl.yaml batch_size=16 num_points=15000 timing=True flops=True
note: 1. set --cfg to cfgs/s3dis to profile all models under the folder. 2. you have to install the latest version of DeepSpeed from source to get a correct measurement of FLOPs

Reference

@inproceedings{armeni2016s3dis,
  title={3d semantic parsing of large-scale indoor spaces},
  author={Armeni, Iro and Sener, Ozan and Zamir, Amir R and Jiang, Helen and Brilakis, Ioannis and Fischer, Martin and Savarese, Silvio},
  booktitle=CVPR,
  pages={1534--1543},
  year={2016}
}