DEPTHOR (ICCV 2025)

📰 This repository contains the official implementation of our paper:
"DEPTHOR: Depth Enhancement from a Practical Light-Weight dToF Sensor and RGB Image".

📄 Paper (arXiv) | 🌐 Project Page

---

Installation

Set up the environment using conda:

conda create -n depthor python=3.8
conda activate depthor
pip install -r requirements.txt

Install bpops from the BPNet (requires CUDA 12.1) for CSPN++ support.
Alternatively, you may use your own CSPN++ implementation and retrain DEPTHOR accordingly.

---

Evaluation

Step 1: Prepare Dataset

Download the ZJU-L5 dataset from Deltar, update the data_path_eval and filenames_file_eval in 'configs/test_zju.txt', the structure of data directory:

└── data
    └── ZJUL5
        ├── data.json
        ├── theater
        │   ├── 1645696174.476698.h5
        │   └── 1645696161.588195.h5
        ├── lab1
        ├── cafe1
        └── cafe2

Step 2: Download Pretrained Models

Place the downloaded checkpoints in the checkpoints/ directory and update the weight_path in 'configs/test_zju.txt':

Model	Checkpoint
Depthor-ZJU-Large	Download
Depthor-ZJU-Small	Download

Step 3: Run Evaluation

python evaluate.py configs/test_zju.txt

---

Training

Step 1: Prepare Dataset

Download the Hypersim dataset and update the dataset path in 'configs/train_hypersim.txt', the structure of data directory:

└── data
    └── evermotion_dataset
        ├── hypersim_train.txt
        ├── hypersim_val.txt
        └── scenes
            ├── ai_001_001
            └── ai_055_010

where the hypersim_train.txt and hypersim_val.txt are provided in 'assets'.

Step 2: Start Training

accelerate launch train_acc.py configs/train_hypersim.txt

---

Acknowledgements

This project builds upon the excellent work of the following open-source repositories:

• Deltar
• BPNet
• Depth-Anything V2
• Hypersim

We thank the authors for their contributions to the community.

---

Citation

If you find this work helpful for your research, please consider citing:

@article{xiang2025depthor,
  title={DEPTHOR: Depth Enhancement from a Practical Light-Weight dToF Sensor and RGB Image},
  author={Xiang, Jijun and Zhu, Xuan and Wang, Xianqi and Wang, Yu and Zhang, Hong and Guo, Fei and Yang, Xin},
  journal={arXiv preprint arXiv:2504.01596},
  year={2025}
}