[ICCV 2025] Consistent Time-of-Flight Depth Denoising via Graph-Informed Geometric Attention

Consistent Time-of-Flight Depth Denoising via Graph-Informed Geometric Attention
Accepted by ICCV 2025 Weida Wang, Changyong He, Jin Zeng, Di Qiu

😆 Highlights

• 🚀 Cross-frame graph fusion for ToF denoising: Fuses motion-invariant graph structures across frames to achieve both temporal consistency and spatial sharpness.

• 🧠 Graph-informed geometric attention (GIGA): Learns graph edges via attention from geometric features, enabling accurate cross-frame correspondence.

• 🔬 Interpretable and robust design: Unrolls MAP optimization with graph Laplacian regularization into a network, achieving high denoising accuracy and generalization to real ToF data.

• 📈 State-of-the-art performance: Outperforms prior works by at least 37.9% MAE and 13.2% TEPE on DVToF; robust on real Kinect v2 data without fine-tuning.

Comparison of denoising accuracy on DVToF dataset (normal / augmented noise)

Environment Setup

Clone this directory

git clone https://github.com/davidweidawang/GIGA-ToF.git

Install Dependencies

Ensure your environment has Python 3.9 or later installed. Use the following command to install required dependencies:

pip install -r requirements.txt

Dataset Preparation

Download DVToF dataset and put it under ./dataset. Extraction code: anwf

Basic Usage

Training

Run the following command to start model training:

python -u train.py \
    -b 32 \
    --dev 2 \
    -lr 2.5e-3 \
    --weight_decay 1e-5 \
    -out "/Path/to/models" \
    -d "/Path/to/results/debug" \
    -in '/Path/to/dataset' \
    -e 400 | tee giga_norm_T_weight.txt

Or you can run the script:

bash train.sh

Inferencing

Run the following command to denoise noisy iq:

python predict.py \
    -in "/Path/to/noise/IQ/dir" \
    -ls "/Path/to/test/list" \
    -out "/Path/to/predicted/IQ/dir" \
    -out_mu "/Path/to/predicted/mu/dir" \
    -m "/Path/to/trained/GIGAToF/model" \

python IQ2corr.py \
   --list_path "/Path/to/test/list"

Or you can run the script:

bash predict.sh

Run Corr2Depth script after completing the above steps"

cd matlab_scripts
matlab
>> CorrToDepth      # Takes a long time

Convert depth from .mat format to .npy. You can also save depth to a .png image by set visualize to True.

python mat2depth.py \
    --list_path "/Path/to/test/list" \
    --visualize False

Evaluation

Evaluate trained model using the command:

python eval.py \
   -in "/Path/to/predicted/depth" \
   -gt "/Path/to/ideal/depth" \
   -flow "/Path/to/optic/flow" \
   -out "/Path/to/result/dir" \
   -list_path "/Path/to/test/list" \
   -v "Version"

Or you can run the script:

bash eval.sh

Remove noise of a single image using a trained GIGA

python predict_single.py \
    -in "demo/raw" \
    -out_iq "./denoised_iq" \
    -m "models/checkpoint_best.pth"

Or you can run the script:

bash predict_single.sh

References

If you find this repository useful for your research, please cite the following work.

@article{wang2025consistent,
  title={Consistent Time-of-Flight Depth Denoising via Graph-Informed Geometric Attention},
  author={Wang, Weida and He, Changyong and Zeng, Jin and Qiu, Di},
  journal={arXiv preprint arXiv:2506.23542},
  year={2025}
}