Multi-Modal RGB-D Scene Recognition Across Domains

PyTorch official implementation of "Multi-Modal RGB-D Scene Recognition Across Domains" in Workshop on Multi-Task Learning in Computer Vision (DeepMTL), ICCV 2021.

We introduce a benchmark testbed for a novel unsupervised domain adaptation problem. We revisited the SUNRGB-D dataset, identifying a subset of scene classes shared among four different 3D cameras with different acquisition properties. Each camera is considered as an RGB-D domain.

Setup

1. Download the complete SUNRGB-D dataset from this link.
2. Use the proposed txt files in the "subsets" folder.

3. To run the model, execute the following code: python train.py
   • Configs DATA_DIR_TRAIN, DATA_DIR_TRAIN_2 and DATA_DIR_VAL must be changed according to the domains
   • With Realsense, FILTER_BEDROOM must be set to True and NUM_CLASSES to 9

Requirements

• Cuda 10.1
• Python 3.7.7
• Torch 1.0.0
• Torchvision 0.2.1
• Other Python Requirements in requirements.txt

Acknowledgement

Code in this repository has been written starting from Translate-to-Recognize

Citation

To cite, please use the following reference:

@inproceedings{FerreriBucciTommasi2021,
  title={Multi-Modal RGB-D Scene Recognition Across Domains},
  author={Andrea Ferreri, Silvia Bucci, Tatiana Tommasi},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
  year={2021}
}