Object detection system using deformable part models (DPMs) and latent SVM (voc-release5). You may want to use the latest tarball on my website. The github code may include code changes that have not been tested as thoroughly and will not necessarily reproduce the results on the website.
Home Page: http://www.cs.berkeley.edu/~rbg/latent/
License: MIT License
Information
===========
Welcome to voc-release5!
This is the companion code-release for my Ph.D. dissertation ("Appendix
C").
Project webpage: http://www.cs.uchicago.edu/~rbg/latent/
Release 5 highlights (see docs/changelog for more details)
* Weak-label structural SVM (wl-ssvm) [4]
* Person grammar model (NIPS 2011) [4]
* Convex optimizer improvements (faster convergence)
* Code cleanup, reorganization, and speed improvements
* Training is done in memory (no more large .dat files on disk!)
* Scale (and location) prior
* Star-cascade included
* Bug fixes
Summary
=======
This is an implementation of our object detection system based on
mixtures of deformable part models. This release extends the system
in [2], and is described in my dissertation [5]. The models in this
implementation are represented using the grammar formalism presented
in [3,4,5]. The learning framework supports both binary latent SVM
and weak-label structural SVM (WL-SSVM), which was introduced in
[4,5]. The code also supports the person object detection grammar
from my NIPS 2011 paper [4].
This distribution contains object detection and model learning code,
as well as pretrained models for the PASCAL{07,10} and INRIA Person
datasets. This release also includes code for rescoring detections
based on contextual information and the star-cascade detection
algorithm of [6].
The system is implemented in MATLAB, with various helper functions
written in MEX C++ for efficiency reasons.
More details, especially about the learning algorithm and model
strcuture, can be found in my dissertation [5].
For questions concerning the code please read the FAQ first
http://people.cs.uchicago.edu/~rbg/latent/voc-release5-faq.html,
and then, if you still have questions, contact Ross Girshick at
<ross.girshick AT gmail DOT com>.
This project has been supported by the National Science Foundation
under Grant No. 0534820, 0746569 and 0811340.
How to Cite
===========
If you use this code or the pretrained models in your research,
please cite [2] and this specific release:
@misc{voc-release5,
author = "Girshick, R. B. and Felzenszwalb, P. F. and McAllester, D.",
title = "Discriminatively Trained Deformable Part Models, Release 5",
howpublished = "http://people.cs.uchicago.edu/~rbg/latent-release5/"
}
If you use an intermediate release downloaded from github you may
also want to cite the date or git commit hash.
You should also cite some of the following depending on what aspects
of this system you are using or comparing against:
* [4] for the NIPS 2011 person grammar model and/or Weak-Label
Structural SVM
* [6] for the cascade detection algorithm
* [5] if you discuss specific parts of the system that are not
published elsewhere (e.g., max regularization, latent
orientation)
References
==========
[1] P. Felzenszwalb, D. McAllester, D. Ramaman.
A Discriminatively Trained, Multiscale, Deformable Part Model.
Proceedings of the IEEE CVPR 2008.
[2] P. Felzenszwalb, R. Girshick, D. McAllester, D. Ramanan
Object Detection with Discriminatively Trained Part Based Models.
IEEE Transactions on Pattern Analysis and Machine Intelligence,
Vol. 32, No. 9, September 2010.
[3] P. Felzenszwalb, D. McAllester
Object Detection Grammars.
University of Chicago, Computer Science TR-2010-02, February 2010
[4] R. Girshick, P. Felzenszwalb, D. McAllester
Object Detection with Grammar Models.
Proceedings of Neural Information Processing Systems (NIPS) 2011.
[5] R. Girshick
From Rigid Templates to Grammars: Object Detection with Structured Models.
Ph.D. dissertation, The University of Chicago, April 2012.
[6] P. Felzenszwalb, R. Girshick, D. McAllester
Cascade Object Detection with Deformable Part Models.
In Proceedings of the IEEE CVPR 2010.
System Requirements
===================
* Linux or OS X
* MATLAB
* GCC >= 4.2 (or an older version if it has OpenMP support)
* At least 4GB of memory (plus an additional ~0.75GB for each
parallel matlab worker)
The software was tested on several versions of Linux and Mac OS X
using MATLAB versions R2011a. There may be compatibility issues
with older versions of MATLAB, though I have successfully run it
back to R2008b.
Getting started
===============
1. Unpack the code
2. Start matlab
3. Run the 'compile' function to compile the helper functions.
(you may need to edit compile.m to use a different convolution
routine depending on your system)
4. Run demo.m or demo_cascade.m to see the detection code in action
Note: If you don't start matlab in the code directory, you may need
to manually run startup.m to ensure all paths are correctly set.
When you run the code for the first time you will likely see this
message:
"""
~~~~~~~~~~~ Hello ~~~~~~~~~~~
voc-release5 is not yet configured for learning.
You can still run demo.m, but please read
the section "Using the learning code" in README.
(Could not find the PASCAL VOC devkit in %s)
"""
This is warning you that you have not configured voc_config.m to
look in the correct path for the PASCAL VOCdevkit. You will need
to download the PASCAL VOCdevkit, unpack it, and set the BASE_DIR
variable in voc_config.m accordingly.
Example detection usage
=======================
>> load VOC2007/car_final.mat; % car model trained on the PASCAL 2007 dataset
>> im = imread('000034.jpg'); % test image
>> bbox = process(im, model, -0.5); % detect objects
>> showboxes(im, bbox); % display results
The main functions defined in the object detection code are:
boxes = imgdetect(im, model, thresh) % detect objects in image im
bbox = bboxpred_get(model.bboxpred, dets, boxes) % bounding box location regression
I = nms(bbox, overlap) % non-maximal suppression
bbox = clipboxes(im, bbox) % clip boxes to image boundary
showboxes(im, boxes) % visualize detections
visualizemodel(model) % visualize models
Their usage is demonstrated in the 'demo' script.
The directories 'VOC20??' contain matlab .mat file with models
trained on several PASCAL datasets (the train+val subsets). Loading
one of these files from within matlab will define a variable 'model'
with the model trained for a particular object category in the
current workspace. The value 'model.thresh' defines a threshold
that can be used in the 'imgdetect' function to obtain a high recall
rate.
Using the learning code
=======================
1. Download and install the 2006-2011 PASCAL VOC devkit and dataset
(you should set VOCopts.testset='test' in VOCinit.m)
The code expects to find the VOCdevkit in the path
<BASE_DIR>/VOC<PASCAL_YEAR>/VOCdevkit
2. Modify 'voc_config.m' so BASE_DIR and PASCAL_YEAR are set
to where you've unpacked the VOCdevkit
3. Start matlab
4. Run the 'compile' function to compile the helper functions
(you may need to edit compile.m to use a different convolution
routine depending on your system)
5. Use the 'pascal' script to train and evaluate a model
example:
>> pascal('bicycle', 3); % train and evaluate a 6 component bicycle model
The learning code saves a number of intermediate models in a model
cache directory defined in 'voc_config.m'.
Context Rescoring
=================
This release includes code for rescoring detections based on contextual
information. Context rescoring is performed by class-specific SVMs.
To train these SVMs, the following steps are required.
1) Models for all 20 PASCAL object classes must be trained.
2) Detections must be computed on the PASCAL trainval and test datasets.
(The function trainval.m can be used for computing detections on the
trainval dataset.)
3) Compile the included libsvm matlab interface:
>> cd external/libsvm-3.12/matlab/
>> libsvm_make
After these steps have been completed, the context rescoring can be
executed by calling 'context_rescore()'.
Example:
>> context_rescore();
Cascaded Detection
==================
The star-cascade algorithm [7] is now included with the rest of
object detection system.
Multicore Support
=================
In addition to multithreaded convolutions (see notes in compile.m),
multicore support is also available through the Matlab Parallel
Computing Toolbox. Various loops (e.g., negative example data mining,
positive latent labeling, and testing) are implemented using the 'parfor'
parallel for-loop construct. To take advantage of the parfor loops,
use the 'matlabpool' command.
example:
>> matlabpool open 8 % start 8 parallel matlab instances
The parfor loops work without any changes when running a single
Matlab instance. Note that due to the use of parfor loops you may
see non-sequential ordering of loop indexes in the terminal output when
training and testing. This is expected behavior. The parallel computing
toolbox has been tested on Linux using Matlab 2011a.
The learning code, which uses Mark Schmidt's minConf for LBGFS with
simple box constraints, now computes function gradients using OMP
based multithreading. By default a single thread is used unless a
matlabpool has already been opened. Note that when computing the
function gradient with different numbers of threads, the resulting
gradients will be very slightly different. In practice this leads
to small variations in the resulting AP scores.
Example training time
=====================
I just trained a 2007 bicycle model on a new 6-core Intel(R) Core(TM)
i7-3930K CPU @ 3.20GHz system using a pool of 8 matlab workers in
71 minutes. Testing on the 2007 test set took an additional 38
minutes.
parsing model
??? Subscript indices must either be real positive integers or logicals.
Error in ==> train at 217
model.thresh = pos_vals(ceil(length(pos_vals)*0.05));
I modified the code for windows and there was no problem. I trained a couple of times successfully but suddenly this run-time error showed up. Exactly after "Cache holds ... examples".
I‘m training by some grayscale images。
Last time I used 70 positive samples and 10 negative samples to train a detectable model。
But this time i use some other grayscale images for training (50:20, 120:20), a lot of NaN will appear during the training process
Does anyone know what is causing this？
I encountered an error when learning from my own data. I think it relates to the parallel computing part.
Error using parallel_function (line 604)
All workers aborted during execution of the parfor loop.
Error in train>poslatent (line 503)
parfor k = 1:thisbatchsize
Error in train (line 113)
= poslatent(t, iter, model, pos, fg_overlap, num_fp);
Error in pascal_train (line 76)
models{i} = train(models{i}, spos{i}, neg_small, false, false, 4, 3, ...
Error in pascal (line 49)
model = pascal_train(cls, n, note);
The client lost connection to worker 1. This might be due to network problems, or the interactive communicating job might have errored.
I am using ubuntu14.04 with matlab2015b.
I did check the log file but didn't figure it out what's the cause. Any ideas? Thanks.
I have met a problem when I run the demo.m, and the error informations are as follows:
///// Running demo for car /////
input image
press any key to continue
continuing...
car model visualization
press any key to continue
continuing...
Invalid MEX-file
'/media/ExtHDD/ygj/EclipseWorkspace/cachedir/voc-dpm-master/bin/resize.mexa64':
/media/ExtHDD/ygj/EclipseWorkspace/cachedir/voc-dpm-master/bin/resize.mexa64:
failed to map segment from shared object
Error in featpyramid (line 56)
scaled = resize(im, 1/sc^(i-1));
Error in imgdetect (line 25)
pyra = featpyramid(im, model);
Error in demo>test (line 67)
[ds, bs] = imgdetect(im, model, -1);
Error in demo (line 27)
test('000034.jpg', model, 1);
My environment:
Does anyone know how to solve it? Thanks in advance.
Hey:
Hi!
I can't find a explanation of the different fields of the Matlab model datatype and its relations, which would be very nice for understanding.
model =
class: 'ExampleModel'
year: '2014'
note: '29.Apr.2014:09.09.13'
filters: [1x108 struct]
rules: {1x222 cell}
symbols: [1x222 struct]
numfilters: 108
numblocks: 228
numsymbols: 222
start: 1
maxsize: [17 7]
minsize: [11 5]
interval: 10
sbin: 8
thresh: 0.9999
type: 'M'
blocks: [1x228 struct]
features: [1x1 struct]
stats: [1x1 struct]
bboxpred: {12x1 cell}
Hey there!
Because I could not find your email adress to contact you directly I am doing it here now:
Could you please state something about the relationship between voc-dpm and http://docs.opencv.org/master/modules/objdetect/doc/latent_svm.html? Maybe in the Wiki or Readme here or at OpenCV.
As far as I see, you can not train a model with the software included in OpenCV. This is why I came here. Am I right?
Thanks for your great work!
I am using macOS Mojave with MATLAB_R2018b. Is there a way I can use a different gcc version than the Xcode one?
In the Readme (https://github.com/rbgirshick/voc-dpm/blob/master/README#L182) it says:
Using the learning code
- Download and install the 2006-2011 PASCAL VOC devkit and dataset
(you should set VOCopts.testset='test' in VOCinit.m)
I dont understand that suggestion as there is no set named 'test' in the PASCAL datasets, only train, val und train val.
So, what does it mean? Specifing a self-made test-set?
Matlab is too huge for my computer......, also projects are based on Python
But may difficult
I encountered the error .Can you please help me out .The line was in
"Error in featpyramid (line 30)
scaled = resize(im, 1/sc^(i-1));
"
Hello, I want to try voc-realse 5.0, but when I run "pascal('bicycle', 3);", I encounter the following error:
Caching features
cluster objective: 6285.721
cluster objective: 6285.721
cluster iter: 23/25
Invalid MEX-file
'/home/jeremy/jWork/voc-dpm-master/bin/fv_cache.mexa64': dlopen:
cannot load any more object with static TLS
Error in train (line 76)
fv_cache('init', max_num, max_dim, max_nbls);
Error in pascal_train (line 71)
models{i} = train(models{i}, spos{i}(inds), neg_large, true,
true, 1, 1, ...
Error in pascal (line 62)
model = pascal_train(cls, n, note);
can someone help me?
http://people.cs.uchicago.edu/~rbg/INRIA_PASCAL.tgz
This link has failed. I really need the annotation file
Hi,
I am a little confused in the process of training:
1. What's the meaning of "belief" in your code?
2. Does ex mean a picture and does fv mean one object in that picture?
3. Where's the code for optimizing param z of negative? And where's the code for training best location z?
4. What's the role of using the loss pyramid playing in your code?
Thank you for your answer. I spend a lot of time to read your code, and I really need some clues. Can you answer in detail？
best wishes,
MX
Hi,
I am working on a project "Tiger Detection with Body Part Localization".
So I want to detect and localize body parts of tiger using Deformable Part Models.
I read your ReadMe file in which you have only discussed how to train and learn a model using Pascal Dataset. Can you please tell me how can I use your code to learn and model my dataset?
Thanks.
Hello, I don't know so much about MEX compilation, excuse the naiveness of my issue
I tried to compile the toolkit under linux x64 and matlab R2014a.
The first error I encountered was that the -o arguments in MEX calls from compile.m were not valid mex arguments. I replaced them by -output arguments, which worked.
Then I got another error : in verbose mode, I get
/usr/bin/g++ -c -DMX_COMPAT_32 -D_GNU_SOURCE -DMATLAB_MEX_FILE -I"/usr/local/MATLAB/R2014a/extern/include" -I"/usr/local/MATLAB/R2014a/simulink/include" \"\-ansi -fexceptions -fPIC -fno-omit-frame-pointer -pthread -Wall -fopenmp" -Iexternal -DMETA_NUM_FEATURES=8 -O3 -DNDEBUG /home/maxim/Documents/voc-dpm-master/gdetect/fconv_sse_meta.cc -o /tmp/mex_299675772102824_29533/fconv_sse_meta.o
/tmp/mex_299675815106937_29533: 3: /tmp/mex_299675815106937_29533: Syntax error: Unterminated quoted string
I could not easily investigate the contents of this /tmp/mex_299675815106937_29533 file because it was deleted upon termination of the mex compilation. I finally ran an infinite loop that copied the mex temporary files in another folder in parallel to inspect its contents:
#!/bin/sh
/usr/bin/g++ -c -DMX_COMPAT_32 -D_GNU_SOURCE -DMATLAB_MEX_FILE -I"/usr/local/MATLAB/R2014a/extern/include" -I"/usr/local/MATLAB/R2014a/simulink/include" \"\-ansi -fexceptions -fPIC -fno-omit-frame-pointer -pthread -Wall -fopenmp" -Iexternal -DMETA_NUM_FEATURES=8 -O3 -DNDEBUG /home/maxim/Documents/voc-dpm-master/gdetect/fconv_sse_meta.cc -o /tmp/mex_299675772102824_29533/fconv_sse_meta.o > /tmp/mex_299675815124147_29533 2> /tmp/mex_299675815135458_29533Can somebody help me understand what is wrong with this file, how and where to fix it ?
Many thanks
(PS: I also posted on mathworks forums http://www.mathworks.fr/matlabcentral/answers/130370-how-can-i-debug-this-failed-mex-compilation)
When running your compile.m, I see this warning
Warning: Maybe you need to call fv_cache('unlock') first?
In fv_compile (line 52)
In compile (line 48)
which means I can not successfully compile fv_cache.cc function.
I saw your comment in this exception, but I don't know how to solve this problem.
There is one guy in stack overflow suggested that he has met the same problem, and he solve it by installing gcc48.
http://stackoverflow.com/questions/24150718/voc-release-5-demo-m-is-not-working/33789824#33789824
Is that the correct way?
systems
Ubuntu
Matlab r2020b
gcc 9.3
While running the compile.m following error shows
Building with 'g++'.
/home/nasim/Desktop/voc-release5/2012/voc-dpm/fv_cache/fv_cache.cc: In function ‘void gradient_handler(int, mxArray**, int, const mxArray**)’:
/home/nasim/Desktop/voc-release5/2012/voc-dpm/fv_cache/fv_cache.cc:417:21: warning: narrowing conversion of ‘dim’ from ‘int’ to ‘mwSize’ {aka ‘long unsigned int’} [-Wnarrowing]
417 | mwSize dims[] = { dim };
| ^~~
/home/nasim/Desktop/voc-release5/2012/voc-dpm/fv_cache/fv_cache.cc: In function ‘void get_model_handler(int, mxArray**, int, const mxArray**)’:
/home/nasim/Desktop/voc-release5/2012/voc-dpm/fv_cache/fv_cache.cc:510:23: warning: narrowing conversion of ‘(int)M.model::num_blocks’ from ‘int’ to ‘mwSize’ {aka ‘long unsigned int’} [-Wnarrowing]
510 | mwSize dims[] = { M.num_blocks };
| ~~^~~~~~~~~~
/home/nasim/Desktop/voc-release5/2012/voc-dpm/fv_cache/fv_cache.cc:515:38: warning: narrowing conversion of ‘(((int)M.model::block_sizes) + ((sizetype)(((long unsigned int)i) * 4)))’ from ‘int’ to ‘mwSize’ {aka ‘long unsigned int’} [-Wnarrowing]
515 | mwSize dims[] = { M.block_sizes[i] };
| ~~~~~~~~~~~~~~~^
MEX completed successfully.
Building with 'g++'.
Error using mex
/home/nasim/Desktop/voc-release5/2012/voc-dpm/star-cascade/cascade.cc: In function ‘double conv(int, int,
const float*, const mwSize*, const float*, const mwSize*, int)’:
/home/nasim/Desktop/voc-release5/2012/voc-dpm/star-cascade/cascade.cc:62:25: warning: comparison of integer
expressions of different signedness: ‘int’ and ‘const mwSize’ {aka ‘const long unsigned int’}
[-Wsign-compare]
62 | for (int xp = 0; xp < B_dims[1]; xp++) {
| ~~~^~~~~~~~~~~
/home/nasim/Desktop/voc-release5/2012/voc-dpm/star-cascade/cascade.cc: In function ‘double rconv(int, int,
int, int, int)’:
/home/nasim/Desktop/voc-release5/2012/voc-dpm/star-cascade/cascade.cc:97:43: error: cannot convert ‘int*’
to ‘const mwSize*’ {aka ‘const long unsigned int*’} in initialization
97 | const mwSize A_dims = MODEL->featdims[L];
| ~~~~~~~~~~~~~~~~~^
| |
| int
/home/nasim/Desktop/voc-release5/2012/voc-dpm/star-cascade/cascade.cc: In function ‘double pconvdt(int,
int, int, int, int, int, int, int, int, int, double)’:
/home/nasim/Desktop/voc-release5/2012/voc-dpm/star-cascade/cascade.cc:116:43: error: cannot convert ‘int*’
to ‘const mwSize*’ {aka ‘const long unsigned int*’} in initialization
116 | const mwSize A_dims = MODEL->featdims[L];
| ~~~~~~~~~~~~~~~~~^
| |
| int
/home/nasim/Desktop/voc-release5/2012/voc-dpm/star-cascade/cascade.cc: In function ‘double partscore(int,
int, int, int, int, int, double)’:
/home/nasim/Desktop/voc-release5/2012/voc-dpm/star-cascade/cascade.cc:195:43: error: cannot convert ‘int*’
to ‘const mwSize*’ {aka ‘const long unsigned int*’} in initialization
195 | const mwSize A_dims = MODEL->featdims[L];
| ~~~~~~~~~~~~~~~~~^
| |
| int
Error in cascade_compile (line 55)
eval(mexcmd);
Error in compile (line 44)
cascade_compile(opt, verb);
The following link is not available anymore, did you move the project somewhere else?
http://www.cs.uchicago.edu/~rbg/latent/
I have trained an myself cascade model using author's pca.mat. But i think it is unreasonable, how can i generate myself pca.mat.
Hello, sorry if it's lame issue, but when I try to compile or make, I got an error using mex.
Error in cascade_compile (line 55)
eval(mexcmd)
Error in compile (line 64)
cascade_compile(opt,verb);
I really don't know what to do to get the demo working i would love id you can help me.
Im getting the error "matlab/voc-dpm/gdetect/fconv_sse_meta.cc:22:10: fatal error:
'boost/preprocessor/repeat.hpp' file not found
I see the file is there so I am curious if there is something in the setup on the Mac that I am missing.
Hi everyone,
I successfully trained a model on my own dataset, but when try to uses cascade detection I got an error about pca score files. How can i generate socres for my dataset to use in cascade detection?
After getting compile the files, the demo run provides this output shown below:
demo
input image
press any key to continue
continuing...
car model visualization
press any key to continue
continuing...
Error using resize
Requested 480x640x257612297838540 (17179869184.0GB) array exceeds maximum array size
preference. Creation of arrays greater than this limit may take a long time and cause
MATLAB to become unresponsive. See array size limit or preference panel for more
information.
Error in featpyramid (line 31)
scaled = resize(im, 1/sc^(i-1));
Error in imgdetect (line 15)
pyra = featpyramid(input, model);
Error in demo>test (line 31)
[dets, boxes] = imgdetect(im, model, -0.3);
Error in demo (line 4)
test('000034.jpg', model);
I am trying to run the compile script on a windows machine , with MATLAB 2015.
However, I am facing the following error with the compile.m script:
>> compile
Building with 'Microsoft Visual C++ 2013 Professional'.
Error using mex
resize.cc
C:\Users\Pallavi\Documents\Khimya\voc-release3.1\resize.cc(36) : error C2057: expected constant expression
C:\Users\Pallavi\Documents\Khimya\voc-release3.1\resize.cc(36) : error C2466: cannot allocate an array of constant size 0
C:\Users\Pallavi\Documents\Khimya\voc-release3.1\resize.cc(36) : error C2133: 'ofs' : unknown size
C:\Users\Pallavi\Documents\Khimya\voc-release3.1\resize.cc(70) : error C3861: 'bzero': identifier not found
Error in compile (line 1)
mex -O resize.cc
I did follow the read me and tried all three convolution options as mentioned in step 3
"(you may need to edit compile.m to use a different convolution routine depending on your system) "
Could you please suggest if this code would work in windows ? Also, how could one fix the above issue ? It will be of great help. Thanks.
Even though the testing algorithm in the demo works perfectly for some images, for some others it's not working well and produces an error.
In particular, the error is in reduceboxes function (in util folder). In some images the size(bs)=[0,0] so it cannot access those.
I couldn't compile the project due to following error.
Error using mex
voc-release5/fconv not found; check that you are in the correct current folder, and check the spelling of '/.../voc-release5/fconv'
Error in compile (line 64)
eval([mexcmd 'gdetect/fconvsse.cc -o fconv']);
I am using matlabr2015a on ubuntu 15.10.
Thanks.
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.