pointnet2_rsmix

RSMix for PointNet++(TensorFlow)

We utilize the original released codes of PointNet++, which is implemented with TensorFlow.

• RSMix is implemented in rsmix_provider.py.

Prepare Dataset

• You can get sampled point clouds of ModelNet40 (XYZ and normal from mesh, 10k points per shape) here (1.6GB).

Move the uncompressed data folder or create symbolic link to data/modelnet40_normal_resampled.

• You can also get sampled point clouds of ModelNet40 (XYZ and normal from mesh, 10k points per shape) as hdf5 format here (435MB).

Move the uncompressed data folder or create symbolic link to data/modelnet40_ply_hdf5_2048.

Environment

Follow the environment setting of original PointNet++ code.

[PointNet++ Environment setting]

Point Cloud Classification

Training

Run the training script:

python train.py --model pointnet2_cls_ssg --log_dir ./log/$exp_name $conventional_data_augmentation_arguements $dataset_arguments --beta 1.0

--> --beta 1.0 : RSMix argument

• conventional data agumentation arguments:

(if you use these arguments, input the argument --convda in advance)

--shuffle : Random shuffle augmentation

--jitter : Jitter augmentation

--rot : Random Rotation augmentation

--rdscale : Random Scaling augmentation

--shift : Random Shif augmentation

• RandDrop augmentation argument:

--rddrop : RandDrop augmentation

• Dataset_argument:

(Default dataset: modelnet40)

--modelnet10: modelnet10 argument

• For PointNet training:

replace the model name as pointnet_cls_rsmix

Evaluation

Run the evaluation script:

PointNet++

For ModelNet40 evaluation,

• Single-view evaluation:

python evaluate.py --num_votes 1 --model_path ./log/$exp_name/model.ckpt

• Multi-view evaluation:

python evaluate.py --num_votes 12 --model_path ./log/$exp_name/model.ckpt

For ModelNet10 evaluation,

• Single-view evaluation:

python evaluate_modelnet10.py --num_votes 1 --model_path ./log/$exp_name/model.ckpt

• Multi-view evaluation:

python evaluate_modelnet10.py --num_votes 12 --model_path ./log/$exp_name/model.ckpt

PointNet

For ModelNet40 Evaluation,

• Single-view evaluation:

python evaluate.py --num_votes 1 --model pointnet_cls_basic --model_path ./log/$exp_name/model.ckpt

• Multi-view evaluation:

python evaluate.py --num_votes 12 --model pointnet_cls_basic --model_path ./log/$exp_name/model.ckpt

Save and Visualize Samples

You can input additional arguments related to the other augmentations if you want.

• Save mixed samples:

python train_data_mix_save.py  --model pointnet2_cls_ssg --log_dir $log_dir --mixed_data_dir ./data_mix --mixed_data_save --beta 1.0 $additional_augmentation_arguments

• Visualize Samples:

python ./utils/show3d_balls_rsmix.py  --background_white --ball_mix(if you  want knn visualize, use --knn_mix) --path $mixed_data 

We utilize the viusalization tool in released code of PointNet++.

The original code for visualization is here