This project contains the codes for our paper "Deeply Exploiting Long-Term View Dependency for 3D Shape Recognition".
- python 3.6 (or higher)
- mxnet 1.5.0 (or higher)
- tqdm
To train the model on GPUs, please make sure CUDA is correctly installed on your device. The codes are well tested with mxnet 1.5.1/1.6.0 under CUDA 9.2.
-
Download ModelNet40 and ModelNet10 from their official site.
-
Render 2d images from all .off files using the matlab codes provided by MVCNN
-
Follow the folder structure to re-arrange the rendered image files:
├── test │ ├── airplane │ │ ├── airplane_0627_0001.jpg │ │ ├── airplane_0627_0002.jpg │ │ ├── ...... │ │ ├── airplane_0627_0012.jpg │ │ ├── ...... │ ├── bathtub │ │ ├── ...... ...... ├── train ......
You can also directly download our preprocessed datasets from here.
python train_models.py --batch_size 4 --batch_update_period 32 --num_views 12 --num_classes 10 --dataset_path /xxxxx/xxxx/modelnet10
Notes: Due to limited GPU memory, we can only feed the network with a small batch size (about 2~4). To prevent unstable training with a small batch size, we manually aggregate the gradient and do back propagation every batch_update_period.
python test.py --batch_size 4 --num_views 12 --num_classes 10 --dataset_path /xxxxx/xxxx/modelnet10 --checkpoint /xxxx/xxxx/xxx.params
To reproduce the results in our paper, use the corresponding checkpoints to evaluate the models. Please download the checkpoint files from here.
if you find our work useful in your research, please considering citing:
@ARTICLE{8794586, author={Y. {Xu} and C. {Zheng} and R. {Xu} and Y. {Quan}}, journal={IEEE Access}, title={Deeply Exploiting Long-Term View Dependency for 3D Shape Recognition}, year={2019}, volume={7}, number={}, pages={111678-111691},}
The repo is released under the MIT License. See the LICENSE file for more details.