By Yu-Xiao Guo, Xin Tong

OS: Ubuntu-16.04,
Python: 3.5,
TensorFlow: 1.3.0-RC2,
CUDA: 8.0,
CUDNN: 6.0,
GPUs: NVidia GTX TITAN XP * 2

1. Install TensorFlow: pip install tensorflow-gpu==1.3.0-rc2
2. Compile custom ops: cd libs && source build.sh
3. Prepare training/test samples:
   • Download SSCNet-SUNCG training/test samples: url. (If someone finds the link is invalid, please ask the permission from the author of SSCNet directly)
   • Run: cd tools && python prepare_data.py. Please set DATA_DIR and RECORD_DIR to your local path in advance.
4. Train: source run_training.sh
5. Test: source run_test.sh

• --input-previous-model-path: model dir/file for fine-tune.
• --input-training-data-path: the dir to folder of training TFRecords
• --input-validation-data-path: the dir to folder of test TFRecords
• --input-gpu-nums: gpu nums for training
• --input-network: network structure to train/test, optional choices including VVNetAE30, VVNetAE60, VVNetAE120. If someone tends to try other models in folder models but fails, please feel free to ping us.
• --max-iters: maximum iterations for training, default 150K
• --record-iters: saving model period per iterations, default 2K
• --batch-per-device: batch size per gpu, default 2
• --output-model-path: the dir to save trained models
• --log-dir: the dir to save logs
• --eval-platform: the test output format. fusion will save test tensors with compatible mode with SSCNet evaluation pipeline.
• --eval-results: the folder to save test output
• --phase: the phase of training or test

• tools/prepare_data.py: generating the train/test samples from SSCNet-SUNCG
• refresh_mask.py: generating the mask used by loss function to select proper voxel behind the surface
• tools/mask2cube.py: visualize the mask file saved by refresh_mask.py

Please cite our work if you find helpful in your research:

@InProceedings{guo2018view,
  author={Guo, Yu-Xiao and Tong, Xin},
  title={View-volume network for semantic scene completion from a single depth image},
  booktitle = {IJCAI},
  year={2018}
}