Meowuu7 / Point-Transformer

Pytorch Implementation of Point Transformer

• Implementation of the Point Transformer in Pytorch.
• Some code are borrowed from Pointnet++, point-transformer and point-transformer-pytorch.

Classification

Data Preparation

• Download ModelNet40 and save in data/modelnet40_normal_resampled.
• Download Stanford 3D Indoor Spaces Dataset (S3DIS) dataset and save in datasets/indoor3d_sem_seg_hdf5_data. (The dataset can also be downloaded automatically when running the code.)

Run

• Please use the following command to train and evaluate point-transformer on ModelNet40 for shape classification task:

  python3 main_training.py --num_feat=${num_feat} \ # number of features used for points in the shape cloud
  --device=${device} \ # which gpu to use in the training process
  --batch_size=${batch_size} \ # batch size
  --dp_ratio=${dropout_ratio} \ # dropout ratio
  --attn_mult=${attn_mult} \ # for point-transformer layer
  --task=cls \ # set task to shape classification
  [--use_sgd] \ # use SGD or Adam optimizer
  [--use_abs_pos] \ # whether to use absolute position information in point-transformer layer
  [--with_normal] \ # whether to use points' normal information in point-transformer layer
  [--more_aug] # whether to use more data augmentation in training process

• Please use the following command to train and evaluate point-transformer on S3IDS for semantic segmentation task:

  python3 main_training.py --num_feat=9 \ # number of features used for points in the shape cloud
  --device=${device} \ # which gpu to use in the training process
  --batch_size=${batch_size} \ # batch size
  --dp_ratio=${dropout_ratio} \ # dropout ratio
  --attn_mult=${attn_mult} \ # for point-transformer layer
  --task=sem_seg \ # set task to semantic segmentation
  [--use_sgd] \ # use SGD or Adam optimizer
  [--use_abs_pos] \ # whether to use absolute position information in point-transformer layer
  [--with_normal] \ # whether to use points' normal information in point-transformer layer
  [--more_aug] # whether to use more data augmentation in training process

Results

• Use SGD with initial learning rate 0.001 and decay by 0.1 in epoch 120 and 160, train for 200 epochs; batch size is set to 16; dropout lyaers are added between each two fully-connected layers with the dropout ratio set to 0.4, using absolute position information in the point-transformer layers.

• Instance classification accuracy on ModelNet40 are as follows:

  	OA
  Paper	93.7%
  Ours	92.6%

• Semantic segmentation overall accuracy on S3IDS are as follows:

  	OA
  Paper	90.8%
  Ours	86.0%

• (The performance can be probably improved if the dropout ratio, initial learning rate and the learning rate schduler could be further fine-tuned.)

Miscellaneous

We look forward to authors of Point Transformer releasing their official implementation, training pipelines and trained weights.