search

ostadabbas / An-Efficient-3D-Synthetic-Model-Generation-Pipeline

An Efficient 3D Synthetic Model Generation Pipeline for Human Pose Data Augmentation (CVPRW2021)
5 stars 0 forks source link

readme

An Efficient 3D Synthetic Model Generation Pipeline for Human Pose Data Augmentation

This repository is implementation of the paper An Efficient 3D Synthetic Model Generation Pipeline for Human Pose Data Augmentation which is presented at the CVPR 2021 Workshop on Analysis and Modeling of Faces and Gestures (AMFG2021) and also part of Masters thesis by Kathan Vyas and was developed at the Augmented Cognition Lab.

The given pipeline can use scanned point clouds or pairs of RGB-Depth images using any depth camera (recomended: Intel RealSense D435i) and implement two different variants of Iterative Closest Point (ICP) based on availability of color information for the point clouds. The whole process is displayed in figure below:

image

Under the current repository, codes provided help cover the "Data Collection & Preparation" and "Registration & Mesh Creation" sections.

Table of content

Data Collection Process

For the purpose of this project, We provide two ways to collect the input data: 1) RGB-Depth Images 2) Point clouds

We can collect these from a depth camera. The choice of input depends upon the size and quality of depth. If the scanned point clouds are not able to capture depth well enough, we suggest opting to scan RGB-Depth images separately rather than directly scanning point clouds. More on the collection is provided within Extraction-Setup section. The main algorithm accepts point clouds in either .pcd or .ply format and it can accept any number of point clouds from 2 to 5000. The numbers of point clouds (scans) is something that depends on requirement of the final registered point cloud resolution.

Hardware

In our experiments while creating the pipeline, we used Intel RealSense D435i which is a depth sensing camera from Intel. The Intel RealSense™ D4xx depth cameras can stream live depth (i.e. ranging data) and color data at up to 90 frames per second, and all the processing to generate the depth data is done on-board by the embedded D4 application-specific integrated circuits (ASIC). All the Intel RealSense cameras can be connected to the machine using an USB-C type cable and could be operated using dedicated RealSense SDK. The sdk is only available for windows 10. The SDK provides resources to collect a sequence of RGB-Depth images as well as point clouds.

Extraction-setup

The subject that is being scanned is to be kept in front of camera. You can use a rotating table on which you can put the subject and rotate it. You can then use the RealSense SDK-viewer API which is as shown in figure here.

image

We use the 'record' function on the SDK-viewer API. It records a "bag" file which is a collection of point clouds with time synchronisation.A "BAG" file is nothing but a collection of continuously recorded point clouds. You can record the file within a set frames per second (FPS). This FPS is important because later on when you extract frames or point clouds from the "BAG" file, the number of frames extracted will be equal to fps*seconds_of_recorded_file. We can then either extract point clouds or RGB-D image sequences using Intel RealSense API module [rs-convert] (https://github.com/IntelRealSense/librealsense/tree/master/tools/convert). We suggest to check a few point clouds first to see the quality of depth. If it is good, move to method 1, if not move to method 2.

Method 1. Extraction of Point clouds: Once we have a "BAG" file, you can use the [rs-convert] (https://github.com/IntelRealSense/librealsense/tree/master/tools/convert). The rs-convert is a console app for converting bag files to various formats (currently supported: PNG, RAW, CSV, PLY, BIN). I the observed depth quality in point cloud is good, we can extract point clouds from the bag file. The extracted point clouds could be saved as .ply or .pcd. The code to extract a 10 second "file.bag" which was recorded on 30fps will produce 300 frames using following code:

cd path_to_ Intel RealSense SDK 2.0/tools
.\rs-convert.exe -l path_of_output_folder -i file.bag

Here the parameter "-l" informs rs-convert to extract point clouds in .ply format. Please save the point clouds starting from 0 to n point clouds. The structure of the saved point clouds should be saved as:

folder
    -0.pcd
    -1.pcd
     .
     .
     .
     .
    -n.pcd

Method 2. Extract RGB-Depth image sequence: This is supposed to be used when the depth of the scanned point clouds is of low quality. So we suggest you can scan few point clouds and visualise it first. If the quality is not good, we can extract RGB and Depth images separately. One way to do this is to use [rs-convert] (https://github.com/IntelRealSense/librealsense/tree/master/tools/convert) and execute the code given below. For a 10 second "file.bag" which was recorded on 30fps will produce 300 RGB and 300 Depth frames individually.

For RGB:

cd path_to_ Intel RealSense SDK 2.0/tools
.\rs-convert.exe -c path_of_output_folder -i file.bag

For Depth:

cd path_to_ Intel RealSense SDK 2.0/tools
.\rs-convert.exe -d path_of_output_folder -i file.bag

Here the parameter "-c" informs rs-convert to extract RGB frames only and parameter "-d" informs rs-convert to extract depth frames only. The dimensions of the RGB and Depth images might differ. if so, the next steps would be to bring the dimensions of the depth and RGB same without losing information.

For both RGB and depth (if extracting RGB and depth individually is not supported):

cd path_to_ Intel RealSense SDK 2.0/tools
.\rs-convert.exe -p path_of_output_folder -i file.bag

This will extract all color and depth frames as PNG, we can change them to JPG later. Next, please save the color and depth frames in following structure: 

Data
    images
        -0.jpg
        -1.jpg
        .
        .
        .
        .
        -n.jpg
   Depth
        -0.jpg
        -1.jpg
        .
        .
        .
        .
        -n.jpg

Once this structure is setup, we next move to integrate these RGB and Depth frames together to then form point clouds. Each single pair of RGb and Depth with produce one point cloud. This point cloud would be of better quality then one directly extracted from "bag" file (method 1). For obtaining the point clouds, we use the rgbd_registration.py file.

Installation

The algorithm is purely working in python 3.6 or 3.7. The requirement of the algorithm needs some libraries which can be installed using the ''requirement.txt'' file. The versions of each package are controlled due to Open3D not being backwards compatible as of yet. Please make an environment in python 3.6 or 3.7 using this requirement file and your are ready to go. If you have issues with numpy and open3D, please install numpy firls either using whl or source and then install open3D.

Usage

The code is very self-explanatory and easy to use. There are two main files:

  1. General_ICP.ipynb -> Used if color information is not available
  2. Color_ICP.ipynb -> Used if color information is available

Execution

The execution of each code asks for following argument:

  1. Number of point clouds: This is the total number of point cloud you are using as input. this number as explained earlier can be anywhere between 2 to n.

  2. Input Folder (not asked in this version but will be put as an argument in later version): This is not asked as argument in color_ICP. you can change the directory for the input from each of the file in this segment: image.

  3. Output Folder (not asked in this version but will be put as an argument in later version): This is where the output registered point cloud will be stored. Present version stores a final registered point cloud in same directory as input directory.

Code

This section will look into parts of code responsible for operation of registration. It explains all functions used in the code

  1. For General_ICP.ipynb

These functions are responsible for import of the point clouds. The "load_point_clouds" takes in the input directory, voxel size (one of the hyper-parameters set by user), number of point clouds in the input directory and the angles (if known). The "preprocess_point_cloud" takes in each point cloud from the input directory and calculates global geometric features "Fast Point Feature Histogram" as well as "Normals" for each point in the point cloud. More details about this could be obtained from the paper in subsequent sections.

def load_point_clouds(directory ,voxel_size ,n,g)

def preprocess_point_cloud(pcd, voxel_size)

The "execute_global_registration" implements the RANSAC based global registration. It uses the the calculated geometric features along with pairs of point clouds (source and target) to obtain initial allignment with a RANSAC based inler identification technique. The outpus of this function will be initial alignment for the source and target. The function "pairwise_registration" applies the general ICP algorithm using the inital alignment obtained from global_registration function. It will return the final transformation matrix (includes rotation and translation between source and target). The "main_registration" then implements pose graph to obtained a collective registration of all input point cloud to give one single registered point cloud. The "final_fun" is just a collection of these previously defined functions in one single function.

def execute_global_registration(source_down, target_down, source_fpfh, target_fpfh, voxel_size)

def pairwise_registration(source, target, source_fpfh, target_fpfh, voxel_size)

def main_registration(pcds,fpfh, max_correspondence_dist_coarse,max_correspondence_dist_fine, voxel_size)

def final_fun(dirtry, g)

The "draw_registration_result" allows visualization of the point clouds. It can implement transformation to a point cloud. This function can be used to visualise two point clouds with a without trnsformation applied.

def draw_registration_result(source, target, transformation)
  1. For Color_ICP.ipynb

The only major difference here happens in function presented here below:

def main_registration(pcds,fpfh, max_correspondence_dist_coarse,max_correspondence_dist_fine, voxel_size)

Here we look at three channels in form of scales (primiarily this will be R-G-B channel) and perform scale based color ICP.

Result

The result of the whole process is a final registered point cloid. If you have not made any changes in teh directory of the output file, the final file will be saved in teh same directory as input files. You can change the extension of the output file to either ".ply" or ".pcd" formats. Open3D library is needed for this.The resulting point cloud can be viewed using following code:

import open3d as o3d

pcd = o3d.io.read_point_cloud("../point_cloud.ply")    #point cloud directory                                           #
print(np.asarray(pcd.points))                          #print X-Y-Z coordintaes of the points
o3d.visualization.draw_geometries([pcd])               #Draw_geometries helps to visualise the point cloud.

Or it can also be viewed using either online version of meshlab-online or you can download meshlab on local machine. For more detail on how to import a point cloud in meshlab please look at the instructions found here

Post-Processing

Citation

If you use the code or use part of it, we would appreciate if you could support us by citing our work using:

@InProceedings{Vyas_2021_CVPR, author = {Vyas, Kathan and Jiang, Le and Liu, Shuangjun and Ostadabbas, Sarah}, title = {An Efficient 3D Synthetic Model Generation Pipeline for Human Pose Data Augmentation}, booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops}, month = {June}, year = {2021}, pages = {1542-1552} }

License

Acknowledgements

The whole pipeline code is influenced using the code from open3D library. Also the research papers used for the construction are provided in reference section.

Contact:

Kathan Vyas

Sarah Ostadabbas

References

  1. M. A. Fischler and R. C. Bolles, “Random sample consensus: a paradigm for model fittingwith applications to image analysis and automated cartography,”Communications of the ACM,vol. 24, no. 6, pp. 381–395, 1981
  2. S. Rusinkiewicz and M. Levoy, “Efficient variants of the icp algorithm,” inProceedings thirdinternational conference on 3-D digital imaging and modeling. IEEE, 2001, pp. 145–152
  3. Y. Chen and G. Medioni, “Object modelling by registration of multiple range images,”Imageand vision computing, vol. 10, no. 3, pp. 145–155, 1992
  4. J. Park, Q.-Y. Zhou, and V. Koltun, “Colored point cloud registration revisited,” inProceedingsof the IEEE International Conference on Computer Vision, 2017, pp. 143–152
  5. Q.-Y. Zhou, J. Park, and V. Koltun, “Open3D: A modern library for 3D data processing,”arXiv:1801.09847, 2018