DORNet: A Degradation Oriented and Regularized Network for
Blind Depth Super-Resolution
🌟 CVPR 2025 (Oral Presentation) 🌟

Zhengxue Wang^1*, Zhiqiang Yan✉^1*, Jinshan Pan¹, Guangwei Gao², Kai Zhang³, Jian Yang✉¹

^*Equal contribution    ^✉Corresponding author   
¹Nanjing University of Science and Technology   
²Nanjing University of Posts and Telecommunications    ³Nanjing University   

Overview of DORNet. Given $\boldsymbol D_{up}$ as input, the degradation learning first encodes it to produce degradation representations $\boldsymbol {\tilde{D}}$ and $\boldsymbol D $. Then, $\boldsymbol {\tilde{D}}$, $\boldsymbol D $, $\boldsymbol D_{lr} $, and $\boldsymbol I_{r}$ are fed into multiple degradation-oriented feature transformation (DOFT) modules, generating the HR depth $\boldsymbol D_{hr}$. Finally, $\boldsymbol D$ and $\boldsymbol D_{hr}$ are sent to the degradation regularization to obtain $\boldsymbol D_{d}$, which is used as input for the degradation loss $\mathcal L_{deg}$ and the contrastive loss $\mathcal L_{cont}$. The degradation regularization only applies during training and adds no extra overhead in testing.

Dependencies

Python==3.11.5
PyTorch==2.1.0
numpy==1.23.5 
torchvision==0.16.0
scipy==1.11.3
Pillow==10.0.1
tqdm==4.65.0
scikit-image==0.21.0
mmcv-full==1.7.2

Datasets

RGB-D-D

TOFDSR

NYU-v2

Models

Pretrained models can be found in checkpoints.

Training

For the RGB-D-D and NYU-v2 datasets, we use a single GPU to train our DORNet. For the larger TOFDC dataset, we employ multiple GPUs to accelerate training.

DORNet

Train on real-world RGB-D-D
> python train_nyu_rgbdd.py
Train on real-world TOFDSR
> python -m torch.distributed.launch --nproc_per_node 2 train_tofdsr.py --result_root 'experiment/TOFDSR'
Train on synthetic NYU-v2
> python train_nyu_rgbdd.py

DORNet-T

Train on real-world RGB-D-D
> python train_nyu_rgbdd.py --tiny_model
Train on real-world TOFDSR
> python -m torch.distributed.launch --nproc_per_node 2 train_tofdsr.py --result_root 'experiment/TOFDSR_T' --tiny_model
Train on synthetic NYU-v2
> python train_nyu_rgbdd.py --tiny_model

Testing

DORNet

Test on real-world RGB-D-D
> python test_nyu_rgbdd.py
Test on real-world TOFDSR
> python test_tofdsr.py
Test on synthetic NYU-v2
> python test_nyu_rgbdd.py

DORNet-T

Test on real-world RGB-D-D
> python test_nyu_rgbdd.py --tiny_model
Test on real-world TOFDSR
> python test_tofdsr.py --tiny_model
Test on synthetic NYU-v2
> python test_nyu_rgbdd.py --tiny_model

Experiments

Quantitative comparison

Complexity on RGB-D-D (w/o Noisy) tested by a 4090 GPU. A larger circle diameter indicates a higher inference time.

Visual comparison

Visual results on the real-world RGB-D-D dataset (w/o Noise).

Citation

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

@inproceedings{wang2025dornet,
  title={DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution},
  author={Wang, Zhengxue and Yan, Zhiqiang and Pan, Jinshan and Gao, Guangwei and Zhang, Kai and Yang, Jian},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={15813--15822},
  year={2025}
}