This repository contains the TensorFlow (1.2) implementation of the paper "Deep End-to-End Alignment and Refinement for Time-of-Flight RGB-D modules". This implementation has been tested on Ubuntu 14.04. with CUDA 8.0.
To use it, first create your python virtual environment, and install the requirements by
pip install -r requirements.txt
Then compile the cuda operations: in /utils and /utils/ops/warp_by_flow, type in the Terminal
make
Now you should be ready to go. To test the model, use the fullmodel.py in the /fullmodel directory, where you can set the data directory and output directory.
Our trained model is provided here.
You can download the RGB-Depth dataset (~20GB) here. The loader.py file is responsible for loading this dataset.
Update 17 March, 2020: Since the original transient rendering is too large to host on Drive, we provide the original blender&pbrt files in the same Drive folder with code on generating the data.
Note that our data generation follows the protocol from Su et al. Deep End-to-End Time-of-Flight Imaging, CVPR 2018.
This repo contains the source and data files for generating synthetic ToF data.
There are two major parts in the data generation pipeline:
- Produce ground truth depth using Blender with a .blend file.
- Produce transient renderings using pbrt-v3-tof with a (or multiple) .pbrt file(s), and use MatLab to process the renderings as ToF raw correlation measurements or ToF depth images (no plane correction).
Optionally, you may render corresponding color images using Blender or the official version of pbrt.
The major files in repo is organized as follows:
|- blender_utils
|-- export_path.py # export camera locations in Blender's Timeline, should be run inside of Blender #
|-- output_zpass.py # python script for writing ground truth depth #
|-- lighting_multiple_output # further applications for use of python in blender #
|- pbrt-v3-tof
|-- example
|-- batch_run_pbrt.sh # pbrt rendering example #
|- transient_processing
|-- example
|-- transient_to_depth.m # MatLab script for converting transient rendering into ToF correlation measurements and ToF depth images #
|- pbrt_material_augmentation
|-- exanple
|-- output_materail.m # MatLab script for writing material library .pbrt files #
|- scenes # 3D models and camera paths #
Tested under Ubuntu 14.04
- MatLabEXR: https://blog.csdn.net/lqhbupt/article/details/7867697 (in Chinese)
- pbrt-v3-tof: This is a costomized renderer for transient rendering provided by Su et al. https://vccimaging.org/Publications/Su2018EndToEndTOF/, ordinary cmake build should work. For official pbrt, see https://www.pbrt.org/
- Blender 2.7: https://www.blender.org/
- Render ground truth depth images (no plane correction) using the Depth Pass in the Blender's Cycles renderer. In Blender's GUI, it is visible in the Node editor's Renderlayers viewport. Make sure the "Use Nodes" is ticked. Set your camera position and hit the camera-shot button to render the image and save it into the path.
The python script for producing this procedure without using the GUI is given in blender_utils/output_zpass.py. The termnial command is
$blender_path/blender -b $.blendfile --python output_zpass.py --$python_args
Rendering the depth pass is very fast in Blender. You can do more interesting things with blender's python library bpy, as illustrated in the lighting_multiple_output.py.
- Render transient images. This is a one-line command
$pbrt_path/pbrt $pbrt_file
which will produce by default 256 transient images in $pwd. Typically it taks 3~5 minutes to render on a multicore CPU. A batch processing file is given in /pbrt-v3-tof/example/batch_run_pbrt.sh.
- Process the transient images. Here we use MatLab for this purpose. The code is provided in /transient_processing/example/transient_to_depth.m.
It is important that the .blend file and .pbrt file correspond to the same scene at the same camera viewpoint. There are differences between the coodinate system used by pbrt-v3-tof and Blender. For example, if you want to produce a .pbrt file from a .blend file, first use Blender's "Export as .obj" function and choose the option "-Z forward & Y up", and then further transform it into .pbrt using obj2pbrt provided in the pbrt-v3-tof package. You often have to make sure the rendering result are consistent. If they are still not consistent, you may also want to check other parameters such as fov, image resolution, etc. in both Blender's setting and the setting in the .pbrt file. As far as I know, blender refer to the horizontal dimension for FOV while pbrt uses the vertical dimension.
In /pbrt_material_augmentation we have some utility functions on augmenting material properties in .pbrt file using MatLab. What it does is to replace the material parameters with some prescribed or random numbers. You should refer to the format of .pbrt files in https://www.pbrt.org/fileformat-v3.html. Note in particular that a .pbrt file can refer to other .pbrt files. This will come handy if we have a material library .pbrt file as we do assume here.
-
This pipeline is heavily influenced by Su et al.'s repo: https://vccimaging.org/Publications/Su2018EndToEndTOF/. You can find similar working examples and scene files there.
-
The principle of using transient rendering to approximate multi-path interference error in time-of-flight image is described in many places. See e.g.
-
I have a exposition about the principle of path tracing in here. Transient rendering is briefly touched. For full reference, please go to https://www.pbrt.org/, one of the best expositions available online.
-
Blender tutorial: https://www.youtube.com/user/AndrewPPrice
In the Drive folder there is a calib.bin file, which should be used if mvg_aug is set to be true when training. There is another calib.bin file in the test_real folder, which contains a sample real test image. The calib.bin file there is for real calibration.
For the most updated details about this work, please refer to the arxiv paper. If you find this work useful, please cite
@inproceedings{qiu2019rgbd,
title={Deep End-to-End Alignment and Refinement for Time-of-Flight RGB-D modules},
author={Di Qiu, Jiahao Pang, Wenxiu Sun, Chengxi Yang},
booktitle={International Conference in Computer Vision (ICCV)},
year={2019}
}
- Can we possibly find the "correct material"? In this pipeline, given fixed scene geometry, multi-path interference error is determined by the material properties. Determining what meterial specification should coincide with the sensor's capture is not done here.
- Sensor error simulation In Guo's FLAT dataset, they implemented Kinect's camera function and modelled its noise distribution. Here we do not have sensor specific data pipeline.
- GPU acceleration PBRT currently only parallelizes on CPU.
- PBRT-Blender materal correspondence This is specifically important if one wants to use RGB images from Blender.
Disclaimer: This software and related data are published for academic and non-commercial use only.