CMU
15-418 (Spring 2012) Final Project:
Parallel Vision Detection on GPU
Pinjie Huang
Video Demo
Reports
Week |
What
We Plan To Do |
What
We Actually Did |
Apr 1-7 |
Document Study:
HOG, latent SVM, 3-D convolution |
Document Study |
Apr 8-14 |
Load image onto
GPU, implement HOG on GPU |
Run the sequential starter
code voc-release4.01 with appropriate setting |
Apr 15-21 |
Implement parallel
training of several online SVMs on GPU |
Implement
feature.cc in CUDA with optimizations |
Week |
What We Plan
To Do |
What We
Actually Did |
Mon
4/23- Sun 4/29 |
Optimized fconv.cc. Implemented convolution of the calculated features
and the model in GPU. |
Written fconv.cc.
Linked and compiled the thing with Matlab in GPU. |
Mon
4/30-Thu 5/3 |
Link the feature calculation and convolution with the original code. |
Used ffmpeg for
decoding and encoding. |
Fri
5/4-Sun 5/6 |
Performance
Analysis. Speed-up of GPU version against CPU version for each part. Portion
of time is consumed. |
Debugging. Run-time
analysis. |
Mon
5/7-Thu 5/10 |
Prepare for Presentation,
Final Report. |
|
Working Log
[4/2/2012]
Project Proposal.
[4/12/2012]
Sequential
vision detection.
[4/14/2012]
Cuda
code for histograms over pixels.
[4/18/2012]
Working
code of feature calculation.
[4/20/2012]
Found
a Guide to Using NVMEX tool : http://www.cs.ucf.edu/~janaka/gpu/using_nvmex.htm.
[4/22/2012]
Checkpoint Report.
[4/23/2012]
[5/1/2012]
Linked
and compiled all the pieces.
[5/7/2012]
GPU
version Code now can detect person with the same result in CPU.
[5/10/2012]
Project
Presentation.