self.introduction

View project on GitHub

Welcome to Ludan’s pages

Profile

15+ years of embedded systems experience. Skilled in software development and integration on embedded Linux and Android. In the past 5 years, I’ve been involved in machine/deep learning accelerated by heterogeneous computing, ideally with GPU (OpenCL, CUDA), FPGA, DSP. Nowadays, my project is to research and develop unsupervised learning for defect inspection.

Honors

Work Experience

  • HON HAI/Foxconn Technology Group (09/2019 - Present)

  • LIPS Technology (02/2016 - 09/2019)

    Led a team of 6 to carry out several projects for proof of concept and mass production:


    Project : 3D Hand Pose lower latency and higher accuracy compared to Leap Motion.

    To bring your AR/VR experience to the next level, we are developing a CNN that smooths hand and finger tracking looks and feels more organic than ever. We organized every process, including a team for label engineering, creating 3D label tools and training a regression CNN model with fast and accurate design.

    3D Body Pose Real-time detection using TensorRT and OpenVINO

    We created a system to estimate and track 3D poses of multiple persons by using a CNN for 2D pose estimation and extending the resulting skeletons to 3D by means of the sensor depth in calibrated RGB-Depth camera. To be a real time application, we designed and compressed the CNN model as well as accelerated it by using NVIDIA TensorRT and Intel OpenVINO.

    Project : TOF Camera

    Accomplished a variety of the filters and algorithms for the depth calculation, image calibration , noise reduction and CNN-based inference through CUDA/OpenCL on the several architectures including (1). NVIDIA TK1/TX1(CUDA), (2). Mali, Adruino, Intel(OpenCL with SIMD instructions), (3). Altera Cyclone V FPGA(OpenCL in a pipelined fashion), (4). CEVA DSP(intrinsic functions) .

  • HON HAI/Foxconn Technology Group (07/2014 - 02/2016)

    Led a team of 3 to create proof of concept projects

    Project : Facial recognition

    Accuracy rate of 95.75% on the well-known and challenging LFW benchmark. To evaluate facial recognition, We not only successfully trained DeepID2 network via NVIDIA DIGITS but also developed a prototype of facial recognition system composed of face detection(HOG + SVM), alignment, recognition(DeepID2 + cosine similarity).

    Project : Hand gesture

    Increased 200% efficiency for object detection by using CUDA to accelerate the filters. Created the hand gesture system on NVIDIA TK1. We implemented fast R-CNN involving the region proposal(selective search), CNN(CaffeNet network), and recognition(fine tuning the softmax classifier). Furthermore, we invented an algorithm that can efficiently analysis the hand motions under depth images.

  • ALPHA Networks Inc (09/2010 - 07/20147)

    Project : WD TV

    1. BSP integration :
      • Decreased 50% workload by negotiating with chip vendors to solve the parts of issues.
      • Supervised an Android BSP team of 2 in integrating BSP applications and cooperating with chip vendors.
      • Ported existing software to Android platform (implemented services with android binder).
      • Implemented a Linux device driver using GPIO to control peripheral devices.
    2. Hacking and Mimicking AirPlay protocol :
      • Successfully imitate AirPlay protocol.
      • Analysed the network packet of AirPlay protocol (using Wireshark).
      • Mimicked AirPlay protocol and implemented displaying photos, slideshows and videos functions on our box.
    3. Framework development :
      • Improved work efficiency over 300% by using Python to develop back-end services.
      • Designed and created a framework using the D-Bus libraries to achieve greater capability in communication between various program languages of Javascript, C/C++ and Python.
      • Led a team of 2 to design and implement applications UI of MLB and the App Store (using HTML5/JavaScript/css).
      • Designed and implemented the App Store both on the back-end service and the server (using python).
    4. Implemented Web applications UI :
      • (a). Accuweather. (b). Shoutcast. (c). Wizard. (d). Yahoo weather. Fetching the data from specific URL and parsing the returned files (XML or Jason format) to display the information on the screen according to its design specification.

  • YA HORG Electronic Corp (02/2006 - 03/2007)

    Project : DVJ Player

    1. Firmware :
      • Increased 20% efficiency by avoiding unnecessary transactions.
      • Ported IC firmware of the network chip (CS8900, RTL8019) to DSP chip.
      • Created IC firmware of the IDE ATA/ATAPI and I2C on DSP chip.
    2. Software :
      • Successfully implemented playing DVD in reverse.
      • The development of tools of the TFTP client and SHELL command.
      • Developed the UDF file system and implemented the sliding window of the file cache for sending file data to the DVD player.
      • Designed and implemented a microkernel for the task-switching process.
      • Ported DVD application includes image blending, navigation module, demuxer (parsing the mpeg2 packet to the decoder), command interpreter, and image format translation (RGB to UYVY).

  • MITAC International Corp (06/2005 - 02/2006)

    Project : Mass Rapid Transit - Ticket System (MRT)

    • Designed and implemented interprocess communication using IBM MQ.
    • Multithreaded socket programming for the communication between stations and the server.
    • Database using ORACLE OCCI (C++ embedded SQL).
    • UI design using Borland C++.

EDUCATION

  • National Taipei University of Technology
            Master’s Degree in Computer and Information Sciences.
  • National Yunlin University of Science and Technology
            Bachelor’s Degree in Management Information Systems.