DSC2016.md

視訊訊號處理與深度學習應用

slide

Computer Vision

• character recognizer
• image & vedio
• photosynth
• video reenactment
• auto-car
• object/class recognize
• vedio summarization
• 3D reconstruction
• face detection

relationship to data science

• rich information
• unstructured data
• difficult to indexing & query

Homogeneous Coordinates

• 湊成矩陣，把一切都變矩陣相乘
• 原本 除z 是非線性的 -> 多加一個維度w就變線性了

Filter

• digital color image

   light -> 黑白 -> bayer filter
  

• 2D image filter

• box filter

• sobel edge

• sharpness

Epipolar

• Usage
  • Finding depth

• epipolar constrain -> finding matching point F = Fundamental Matrix

平行運算

• CUDA, openCL, openmp

• share momory (each block have one share momory)

  • need to be synchronisation

• Graphics pipeline

   	input processor 
   	      |
   	geometry stuff 
   	      |
   	  pixed stuff
             |
   accumulated pixel result
  

• is GPU SIMD -> NO

  • divergece!
  • GPU is SIMT(thread)

Framework & Library

Parallel

• MATLAB
• use parallel for
• use gpuarray
• CUDA BLAS
• very fast
• NVIDA thrust library
• c++ STL library for CUDA
• very few lines code for vector manipulation
• fast implementation of parallel primitive

Deep Neural Network

• NVIDIA cuDNN
• very fast
• support popular DN framework
• Caffe
• Tensorflow

openCV tutorial

• histogram equalisation
• canny edge detection
• QR marker detection