Welcome to Xiangfei Meng's Home Page
目前在出海公司,带领技术团队,对 xxx (海外直播 app)的美颜美妆效果负责。团队技术栈涉及图形图像算法、深度学习、C++、OpenGL。
2020~2021年,对 xxx(海外短视频app)的 GAN 模型在 iOS 端的推理效率负责。涉及推理引擎在 iOS 端的 GPU 优化。效率领先于 TFLite、CoreML-Metal、MNN。
2018~2019年,对 xxx(海外短视频app)的服务端超分算法负责。
2015~2018年,以第一名成绩考入北京航空航天大学,取得计算机硕士学位。
2011~2015年,在中国民航大学,取得计算机本科学士学位。
工作/居住地:杭州。
业余爱好:音乐。
近期动向:AI Infra 探索中。
工作理念


动捕设备动辄数十万。该项目旨在利用单摄像头,捕捉鱼的动画,重定向到其他的类鱼卡通角色上。
利用运动目标检测和自适应阈值分割,检测鱼的轮廓边缘。
对轮廓用椭圆傅里叶系数做拟合,用来鲁棒地定位中轴线(medial axis)和鱼身关键点。(至此完成动捕)。
对类鱼的卡通角色,做三角剖分,并用摄像头画面中的鱼的关键点作为控制点,用 MLS 算法驱动网格形变。逐帧实时形变实现鱼动画。(至此完成运动重定向)。
Publication: Xiangfei Meng, Junjun Pan, Hong Qin, Pu Ge. Real-time Fish Animation Generation by Monocular Camera. Computers & Graphics. 71(2018): 55-65. [pdf]
Xiangfei Meng, Junjun Pan, Hong Qin. “Motion Capture and Retargeting of Fish by Monocular Camera.” 2017 International Conference on Cyberworlds (CW 2017), IEEE, Chester, UK, September 20-22, 2017. [pdf][video]
纯兴趣驱动,部分实现了 STL 中的 C++ 容器(vector, list, deque)、迭代器(ordinary iterator, constant iterator, reverse iterator)、适配器(stack, queue, priority_queue)、算法(sort, find)。使用底层内存管理、模板特化和偏特化、函数对象等技术。

纯兴趣驱动,C++实现一套光线追踪渲染器,支持三角形和球形两种形状。支持全局光照(直接光照+间接光照),包含环境光、漫反射、高光、反射、投射、软阴影、渗色。渲染康奈尔盒子如上图。
Implemented a ray tracer as an offline renderer. This renderer is equipped with a parser to analyze model files containing triangles and spheres. The scene is then rendered with global illumination, which includes effects of ambient, diffuse, specular, reflection, refraction, soft shadow and color bleeding. The rendering process is speeded up by OpenMP with thread-level parallelization.

纯兴趣驱动,用 C++ 实现一套 Pascal(类C的面向过程语言)编译器。可将高级语言翻译成 intel-i386 汇编代码执行。编译器支持嵌套函数、递归调用、值传递与引用传递功能。
Implemented a Pascal Compiler using pure C++ without other lexical/syntax tools. The compiler parses Pascal source code and translates it to Intel-i386 assembly code. The assembled executable can run on the Windows operating system. The compiler provides full support for nested function definitions, recursion callings of functions, and the parameter passing either by values or references.
用 C4.5 决策树,自动检测新浪微博上的垃圾账号。10折交叉验证,得到 0.706 F1 Score。
Spammer detection is a typical scenario of the application of classifiers. To detect spammers in social networks, we collected 4109 profiles from Weibo, extracted their features, and delivered them to several classifiers. We implemented a Bayesian classifier and a C4.5 Decision Tree classifier. Bayesian classifier is theoretically optimal as long as the conditional probability density functions are known, and the Decision Tree is independent of dimensions. Through the 10-fold cross-validation, the Bayes Classifier showed 58.3% in recall, 85.4% in precision, and 0.693 in F1-Measure, and the Decision Tree showed 60.1% in recall rate, 85.4% in precision, and 0.706 in F1-Measure.
Publication: 孟祥飞, 徐路, 王思雨. 基于新浪微博的社交网络垃圾用户分析与检测[J]. 科技与创新, 2014(15):125-127. [pdf] 张宇翔, 孙菀, 杨家海, 周达磊, 孟祥飞, 肖春景. 新浪微博反垃圾中特征选择的重要性分析[J]. 通信学报, 2016, 37(8):24-33. [pdf]

运动目标检测是传统 CV 领域的一个基础问题。它可以提供运动目标的有无、位置、速度、轨迹、包围盒等信息,为后续高级语义分析提供支持。
我们实现了基于静态相机的实时运动目标检测。
在一个典型的桌面 PC 上(2.9GHz CPU,4GB内存),对 640x480 图像,处理帧率可达 20fps,满足实时标准。
Moving object tracking is a classical problem of Computer Vision since it can provide essential information on the shape and motion of the foreground objects.
We implemented a real-time moving object tracking system with a static camera.
Estimate trajectories of detected objects by a group of Kalman Filters.
The system is robust to the slight shaking of the camera, and the gradual change of the light. It can reach 20 FPS with 640×480 resolution on my laptop equipped with a 2.9GHz CPU and 4GB memory.
xiangfei_meng@163.com