3D DenseNet(torch version) for ModelNet40 dataset
3D DenseNet is using 3D Convolutional(VolumetricConvolution in torch), Pooling, BatchNormalization layers with 3D kernel. This implements is based on DenseNet and fb.resnet.torch. DenseNet introduced in the paper "Densely Connected Convolutional Networks" (CVPR 2017, Best Paper Award)
See the installation instructions for a step-by-step guide.
train.list
and test.list
file;datadir
variable in examples/run_modelnet40.sh
.See the training recipes for addition examples.
For Modelnet40, just run shell examples/run_modelnet40.sh 0,1
, 0,1
is the GPU ids with multi-GPU supported.
cd examples
./run_modelnet40_h5.sh 0,1
Network | Top-1 error | Top-5 error |
---|---|---|
Voxnet | 13.74 | 1.92 |
DenseNet-20-12 | 12.99 | 2.03 |
DenseNet-30-12 | 12.11 | 1.94 |
DenseNet-30-16 | 11.08 | 1.61 |
DenseNet-40-12 | 11.57 | 1.78 |
This implementation differs from the ResNet paper in a few ways:
3D Convolution: We use the VolumetricConvolution to implement 3D Convolution.