someone-t / photon-efficient

This is the official code for "Photon-Efficient 3D Imaging with A Non-Local Neural Network (ECCV2020)"

Photon-Efficient 3D Imaging with A Non-Local Neural Network

ECCV2020 [Spotlight]

Jiayong Peng, Zhiwei Xiong, Xin Huang, Zheng-Ping Li, Dong Liu, Feihu Xu

Table of Contents

1. Introduction
2. Citation
3. Installation
4. Data generation
5. Training
6. Testing
7. Results

---

1. Introduction

Photon-efficient imaging has enabled a number of applications relying on single-photon sensors that can capture a 3D image with as few as one photon per pixel. In practice, however, measurements of low photon counts are often mixed with heavy background noise, which poses a great challenge for existing computational reconstruction algorithms. In this paper, we first analyze the long-range correlations in both spatial and temporal dimensions of the measurements. Then we propose a non-local neural network for depth reconstruction by exploiting the long-range correlations. The proposed network achieves decent reconstruction fidelity even under photon counts (and signal-to-background ratio, SBR) as low as 1 photon/pixel (and 0.01), which significantly surpasses the state-of-the-art. Moreover, our non-local network trained on simulated data can be well generalized to different real-world imaging systems, which could extend the application scope of photon-efficient imaging in challenging scenarios with a strict limit on optical flux.

2. Citation

If you find the code useful in your research, please consider citing:

3. Installation

A. Environment

The code is developed using python 3.6 and Pytorch 1.0 on Ubuntu 16.04. NVIDIA 1080Ti GPU is required for training and testing.

B. Requirements

pip install -r requirements.txt

4. Data generation

A. generate training and validation data

• download NYU V2 data

  bash ./data_gener/TrainData/download_nyu_dataset.bash

• generate RGB-D and albedo with Matlab 2017b

  matlab -nodesktop -nosplash -r ./data_gener/TrainData/ConvertRGBD

• simulate SPAD measurements with Matlab 2017b

  matlab -nodesktop -nosplash -r ./data_gener/TrainData/SimulateTrainMeasurements

B. generate testing data for Middlebury dataset

• matlab -nodesktop -nosplash -r ./data_gener/TestData/middlebury/SimulateTestMeasurements

C. Indoor and outdoor real-world datasets are provided in

• ./data_gener/TestData/realworld

5. Training

• make train and validation list

  python ./training/util/make_train_lists.py

• edit the corresponding parameters in ./training/config.ini

• run the main train file

  python ./training/main.py

• the training will take about 35 hours on NVIDIA 1080Ti.

6. Testing

• edit the corresponding parameters in ./testing/config.ini

• run the main test file

  python ./testing/main.py

• the test can be conducted on simulated Middlebury dataset, indoor and outdoor real-world dataset by selecting different functions in the ./testing/main.py

7. Results

We provide some of the visualization results here.

• comparisons on Middlebury dataset

• comparisons on outdoor real-world scenes

• comparisons on indoor real-world scenes