In this project, we adapt the well-known attention mechanism to 3D point cloud segmentation and demonstrate its effectiveness in combination with U-Net-inspired architecture and various long-tail recognition (LTR) techniques on the SCANNET200 dataset.

🎉We were voted as the best by 123 classmates and 29 teams through inter-team evaluation!🎉

Our poster:

Our GPU is a single 4090 with 24GB VRAM

cd point-transformer
bash conda_env_setup.sh pt

1. Install torch 1.9.0 corresponds to your cuda version (11.1 in our case)

2. pip install h5py pyyaml sharedarray tensorboardx plyfile

3.  cd point-transformer/lib/pointops
   
    python3 setup.py install

   to install custom kernel functions

Only place the dataset if you plan to train the model.
Place the train and test set in point-transformer/dataset/scannet200
The directory should look like:

point-transfomer/
├─ dataset/
│  ├─ scannet200/
│  │  ├─ train/
│  │  │  ├─ scene0000_00.ply
│  │  │  ├─ ...
│  │  ├─ test/
│  │  │  ├─ scene0500_00.ply
│  │  │  ├─ ...
│  │  ├─ train_split.txt
│  │  ├─ val_split.txt
│  │  ├─ test_split.txt

cd point-transformer
bash train.sh

Since we use five models as ensemble voting classifier, we provide five different weights to download.
Follow the procedure to repreduce the results: \

1. cd point-transformer \
2. For i in [0, 1, 2, 3, 4]:
   a. bash download{i}.sh
   "${1}" is the path to test set folder (e.g., ./dataset/test)
   "${2}" is the path to test.txt (e.g., ./dataset/test.txt)
   "${3}" is the path to output folder (e.g., ./submission)
   b. bash inference.sh $1 $2 $3
   c. Move the output folder (e.g., ./submission) into ../Ensemble/all_submissions
3. cd ../Ensemble
4. Run python3 Ensemble.py
5. The result is in Ensemble/ensemble_results

"${1}" is the path to test set folder (e.g., ./dataset/test)
"${2}" is the path to test.txt (e.g., ./dataset/test.txt)
"${3}" is the path to output folder (e.g., ./submission)

Run:

cd point-transformer
bash download0.sh
bash inference.sh $1 $2 $3