MetaSets: Meta Learning on Point Sets for Generalizable Representations

Datasets

We use ModelNet, ShapeNet and ScanObjectNN as our training and test dataset. In benchmark ModelNet -> ScanObjectNN, we use 11 categories. In benchmark ShapeNet -> ScanObjectNN, we use 9 categories. Download links are shown below.

• ModelNet_11
• ScanObjectNN_11
• ShapeNet_9
• ScanObjectNN_9

Download all these datasets and put them into dir Data. The file tree will be like this:

|--data
   |--modelnet_11
   |--scanobjectnn_11
   |--shapenet_9
   |--scanobjectnn_9

Pretrained Model

We trained model on task ModelNet -> ScanObjectNN and ShapeNet -> ScanObjectNN. Download links are shown below.

• ModelNet -> ScanObjectNN
• ShapeNet -> ScanObjectNN

Download all these models and put them into dir models. The file tree will be like this:

|--models
   |--best_model_model2scan.pt
   |--best_model_shape2scan.pt

After download all models and datasets, and put them into dirs, the whole file tree will be like this:

|--data
|--models
|--data_utils.py
|--dataloader.py
|--model_pointnet_meta.py
|--model_utils_meta.py
|--test.py
|--train.py

Requirements

torch==1.4.0
torchvision==0.5.0
torchmeta
h5py
matplotlib
numpy
Cuda 10.0
Nvidia-driver >=440

Train

Modelnet -> ScanObjectNN

# assemble that 4 gpus are available
python train.py -source modelnet_11 -target scanobjectnn_11 -g 0,1,2,3

ShapeNet -> ScanObjectNN

python train.py -source shapenet_9 -target scanobjectnn_9 -g 0,1,2,3

Evaluate

Modelnet -> ScanObjectNN

python test.py ./data/scanobjectnn_11 models/best_model_model2scan.pt

ShapeNet -> ScanObjectNN

python test.py ./data/scanobjectnn_9 models/best_model_shape2scan.pt

Results

• Benchmark I: ModelNet -> ScanObjectNN 68.63%
• Benchmark II: ShapeNet -> ScanObjectNN 56.75%