This is an unofficial inplementation of RotationNet: Joint Object Categorization and Pose Estimation Using Multiviews from Unsupervised Viewpoints in TensorFlow.
python3.5+TensorFlow(tested on 1.8)opencv
Just clone this repository.
Download multi-view images
$ bash get_full_modelnet_png.sh
Down VGG16 pre-trained models which has been trained on the ILSVRC-2012-CLS image classification dataset from TensorFlow
$ bash get_vgg16_model.sh
If you want to train the model w/o viewpoint prediction
# N is your checkpoint iteration
$ python3 train.py --case 2 --center --lr 0.00001 -no N
If you want to train the model with viewpoint prediction
# N is your checkpoint iteration
$ python3 train.py --case 1 --center --lr 0.00001 -no N
If you want to fine-tune model
# N is your checkpoint iteration
$ python3 train.py --case 2 --center --lr 0.00001 --pretrained N
# N is your checkpoint iteration
$ python3 test.py --case 1 --center --pretrained N
| ModelNet40 | with viewpoint prediction | w/o viewpoint prediction |
|---|---|---|
| Reported | 97.37 | null |
| Reproduced | 96.72 | 97.2024 |