README.md

DuCos: Duality Constrained Depth Super-Resolution via Foundation Model
:star2: ICCV 2025 :star2:

Zhiqiang Yan¹, Zhengxue Wang², Haoye Dong¹, Jun Li², Jian Yang², Gim Hee Lee ¹

¹National University of Singapore   
²Nanjing University of Science and Technology   

Dependencies

Please refer to the requirements.

Datasets

Training split

RGB-D-D

TOFDSR

NYU-v2

Lu & Middlebury

Hypersim

Models

Pretrained models can be found in checkpoints.

Training

python main.py --scale 4 --model_name DuCos --save DuCos_X4 --gpus 0

python main_float.py --model_name DuCos --save DuCos_X2.7 --gpus 0

python main_compress.py --model_name DuCos --save DuCos_compress --gpus 0

python main_RGBDD.py --scale 4 --model_name DuCos --isNosiy 0 --save DuCos_RGBDD --gpus 0

python main_TOFDSR.py --scale 4 --model_name DuCos --isNosiy 0 --save DuCos_TOFDSR --gpus 0

Testing

python test_Sync.py --scale 4 --model_name DuCos --pretrain "./ckpts/x4.pth.tar" --gpu 0

python test_Sync_float.py --model_name DuCos --pretrain "./ckpts/x2.7.pth.tar" --gpu 0

python test_Sync_compress.py --model_name DuCos --pretrain "./ckpts/Compressx8.pth.tar" --gpu 0

python test_RealRGBDD.py --scale 4 --model_name DuCos --isNosiy 0 --pretrain "./ckpts/RealRGBDD.pth.tar" --gpu 0

python test_RealTOFDSR.py --scale 4 --model_name DuCos --isNosiy 0 --pretrain "./ckpts/RealTOFDSR.pth.tar" --gpu 0

Citation

If our method proves to be of any assistance, please consider citing:

@article{yan2025ducos,
  title={DuCos: Duality Constrained Depth Super-Resolution via Foundation Model},
  author={Yan, Zhiqiang and Wang, Zhengxue and Dong, Haoye and Li, Jun and Yang, Jian and Lee, Gim Hee},
  journal={arXiv preprint arXiv:2503.04171},
  year={2025}
}