Home

Awesome

Pytorch RotationNet

This is a pytorch implementation of RotationNet.

Asako Kanezaki, Yasuyuki Matsushita and Yoshifumi Nishida. RotationNet: Joint Object Categorization and Pose Estimation Using Multiviews from Unsupervised Viewpoints. CVPR, accepted, 2018. (pdf) (project)

We used caffe for the CVPR submission. Please see rotationnet repository for more details including how to reproduce the results in our paper.

Training/testing ModelNet dataset

1. Download multi-view images

1-1. Download multi-view images generated in [Su et al. 2015]

$ bash get_modelnet_png.sh

[Su et al. 2015] H. Su, S. Maji, E. Kalogerakis, E. Learned-Miller. Multi-view Convolutional Neural Networks for 3D Shape Recognition. ICCV2015.
This is a subset of ModelNet40.

1-2. Download our multi-view images

$ wget https://data.airc.aist.go.jp/kanezaki.asako/data/modelnet40v2png_ori4.tar; tar xvf modelnet40v2png_ori4.tar

Our BEST results are reported on this dataset.

2. Prepare dataset directories for training

$ bash link_images.sh ./modelnet40v1png ./ModelNet40v1 1  
$ bash link_images.sh ./modelnet40v2png ./ModelNet40_20 2

Or

$ bash link_images.sh ./modelnet40v2png_ori4 ./ModelNet40_20

3. Train your own RotationNet models

3-1. Case (2): Train the model w/o upright orientation (RECOMMENDED)

$ python train_rotationnet.py --pretrained -a alexnet -b 400 --lr 0.01 --epochs 1500 ./ModelNet40_20 | tee log_ModelNet40_20_rotationnet.txt

3-2. Case (1): Train the model with upright orientation

$ python train_rotationnet.py --case 1 --pretrained -a alexnet -b 240 --lr 0.01 --epochs 1500 ./ModelNet40v1 | tee log_ModelNet40v1_rotationnet.txt

Training/testing MIRO dataset

1. Download MIRO dataset (414MB)

$ wget https://data.airc.aist.go.jp/kanezaki.asako/data/MIRO.zip  
$ unzip MIRO.zip

2. Prepare dataset directories for training

$ bash link_images_MIRO.sh ./MIRO ./data_MIRO

3. Train your own RotationNet models

3-1. Case (3): Train the model w/ upright orientation

$ python train_rotationnet.py --case 3 --pretrained -a alexnet -b 480 --lr 0.01 --epochs 1500 ./data_MIRO | tee log_MIRO_160_rotationnet.txt