Translate-to-Recognize Networks

Pytorch implementations of Translate-to-Recognize Networks for RGB-D Scene Recognition (CVPR 2019).

Usage

1. Download Reset18 pre-trained on Places dataset if necessary.
2. Data processing.
   • We use ImageFolder format, i.e., [class1/images.., class2/images..], to store the data, use util.splitimages.py to help change the format if neccessary.
   • Use util.conc_modalities.py to concatenate each paired RGB and depth images to one image for more efficient data loading.
3. Configuration.
   Almost all the settings of experiments are configurable by the files in the config package.
4. Train.
   python train.py or 'bash train.sh' (this will record log to saved file)
5. [New] New branch 'multi-gpu' (more recommended) has been uploaded, making losses calculated on each gpu for better balanced usage of multi gpus. You could use this version using this command: git clone -b multi-gpu https://github.com/ownstyledu/Translate-to-Recognize-Networks.git project_name
6. [New] In multi-gpu brach, we add more loss types in the training, e.g., GAN, pixel2pixel intensity. You could easily add these losses by modifying the config file.

Development Environment

• NVIDIA TITAN XP
• cuda 9.0
• python 3.6.5
• pytorch 0.4.1
• torchvision 0.2.1
• tensorboardX

Citation

Please cite the following paper if you feel this repository useful.

@inproceedings{du2019translate,
  title={Translate-to-Recognize Networks for RGB-D Scene Recognition},
  author={Du, Dapeng and Wang, Limin and Wang, Huiling and Zhao, Kai and Wu, Gangshan},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={11836--11845},
  year={2019}
}