Training

• Files are under train/ directory.

• Training datasets are DeepFocus focal stack subset (RGB-D + Focal stack) and DeepFocus light field subset (RGB-D + Focal stack).

• The training datasets need to be downloaded from DeepFocus repository.

• Under the 'train/' directory, it contains the files for DeepFocus training (Focal stack rendering from RGBD input), our focal stack training (default 0.2 m setting and 0.002 m setting), our light field training, and our transparent or reflective object training.

• We also provide the bash scripts to support full-color training.

• Pretrained checkpoints are provided under the checkpoint/ directory

  Link: https://pan.baidu.com/s/1dV4ql160rF6C90aBsevteg?pwd=2025 code: 2025

• The retraining takes about 10 hours on an NVIDIA RTX 3090 GPU.

• The evaluation results will be in the results/ directory.

Evaluation (Quantitative)

• Files are under eval/ directory.
• Evaluation datasets are DeepFocus focal stack subset (RGB-D + Focal stack) and DeepFocus light field subset (RGB-D + Focal stack).
• The evaluation datasets need to be downloaded from DeepFocus repository.
• Under the 'eval/' directory, it contains the files for DeepFocus evaluation (Focal stack rendering from RGBD input), our focal stack evaluation (default 0.2 m setting and 0.002 m setting), our light field evaluation, and our transparent or reflective object evaluation.
• The Evaluation takes about several minutes on an NVIDIA RTX 3090 GPU.
• The evaluation results will be in the results/ directory.

Testing

• Files are under test/ directory.
• Test datasets are in the rgbd-test-small, and the light field test data are from Inria synthetic light field datasets.
• The Evaluation takes about several seconds/minutes on an NVIDIA RTX 3090 GPU.

Requirements

The code is tested with the following environment:

• Ubuntu Ubuntu 22.04.5 LTS
• Python 3.10+
• PyTorch 2.5.1
• CUDA 12.4
• NumPy 1.26.4

Key dependencies:

torch
torchvision
numpy
scikit-image
opencv-python
matplotlib
imageio
wandb
torchmetrics
DISTS-pytorch
odak

Usage

Testing

The results will be in the results/ directory.

To run inference using a pretrained model:

python test/test_holo_rgb.py
python test/test_trans_holo_rgb.py