Aberration-Aware-Depth-from-Focus

Official code for the paper "Aberration-Aware Depth-from-Focus" (ICCP 2023, TPAMI 2023).

TL;DR: Train a network to represent a lens (4D PSFs). Use the network to simulate aberrated focal stacks. Train depth-from-focus models with aberrated focal stacks. Models can generalize better in real world because it learns under aberrations!

Xinge Yang, Qiang Fu, Mohamed Elhoseiny, Wolfgang Heidrich King Abdullah University of Science and Technology, Saudi Arabia

How to use the code?

This work is a sub-project of our DeepLens project. Here we provide the simplified version for re-implementation.

Warm up

#### aberrated-defocued image simulation ####
python 0_warm_up.py

Train a PSF net for a lens

#### input: (x, y, z, foc_dist), output: PSF at (x, y, z) when the lens is focused to foc_dist ####
python 1_fit_psfnet.py

Simulate aberrated focal stack for depth-from-focus model training

#### simulate focal stack with the same aberration and defocus as capture ####
#### Use AiFNet
python 2_aber_aware_dff_aif.py

#### Use DFVNet
python 2_aber_aware_dff_dfv.py

Data folder

#### Actually you can use any RGB-Depth data pairs, because we render focal stacks by ourselves. ####
- deeplens
- dataset
  - Middlebury2014 # Available at https://vision.middlebury.edu/stereo/data/scenes2014/
    - scene1
    - scene2
    - ...
  - Matterport3D # Available by requesting through https://niessner.github.io/Matterport/
    - train
      - aif
      - depth
    - test
      - aif
      - depth
  - FlyingThings3D_FS # Available at https://github.com/albert100121/AiFDepthNet
    - trian
      - focal_stack
    - val
      - focal_stack

Results

Evaluation on simulated data

Depth map predicted by aberration-aware trained model is more accurate.

Evaluation on real-world data

Depth map predicted by aberration-aware trained model is smoother, with fewer artifacts.

Citations

If our code helps your research or work, please consider citing our paper. The following are BibTeX references:

@ARTICLE{10209238,
  author={Yang, Xinge and Fu, Qiang and Elhoseiny, Mohamed and Heidrich, Wolfgang},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Aberration-Aware Depth-From-Focus}, 
  year={2023},
  volume={},
  number={},
  pages={1-11},
  keywords={Lenses;Optical imaging;Training;Estimation;Computational modeling;Optical sensors;Integrated optics;Depth from focus;optical aberration;point spread function;ray tracing},
  doi={10.1109/TPAMI.2023.3301931}}