Course project for Visual Computing course EE5731R.

Write a program in matlab to estimate depth map from video using this method: Depth Map from Video Sequence.

To make it simple, I only reimplemented:

• Disparity Initialization for video.
• Bundle Optimization for video.

Contain some minor changes.

[0] Place The unzipped Video Dataset(s) at ./video/{$DATASETNAME}/src/{$IMAGENAME}.jpg

• $DATASETNAME = Road, url: http://www.cad.zju.edu.cn/home/gfzhang/projects/videodepth/data/Road.zip
• $DATASETNAME = Lawn, url: http://www.cad.zju.edu.cn/home/gfzhang/projects/videodepth/data/Lawn.rar
• $DATASETNAME = Flower, url: http://www.cad.zju.edu.cn/home/gfzhang/projects/videodepth/data/Flower.rar

[1] Compile GCMex by ./toolbox/GCMex/GCMex_compile.m if necessary.

[2] RUN './demo_disparsity_initialization.m' FOR VIDEO SEQUENCE DISPARSITY INITIALIZATION.

• Outpus: ./result/init/{$DATASETNAME}/test{$ID}.mat # Orginal Outputs
• Outpus: ./result/init/{$DATASETNAME}/test{$ID}.png # Rendered Maps

[3] RUN ./demo_bundle_optimization.m FOR VIDEO SEQUENCE BUNDLE OPTIMIZATION.

• Outpus: ./result/bundle_p{$pass_id}/{$DATASETNAME}/test{$ID}.mat # Orginal Outputs
• Outpus: ./result/bundle_p{$pass_id}/{$DATASETNAME}/test{$ID}.png # Rendered Maps

IMPORTANT: Depend on your computer's permorfance, It may cost 30~60 mins for processing 1 frame with 20 neighborhoods. You can (but not recommend) cut the neighborhood frame number for acceleration (~5mins/frame).

MATLAB 2014B x64:

• With Parallel Computing Toolbox.

UBUNTU 14.04 LTS x64

• Intel i7-5820k, 3.3GHz
• Memory DDR4 2133MHz 64gb (16GBx4)

All three datasets have been processed by a distributed computing cluster. (12 computers, 72 cpu cores intotal)

The results are saved under result.

It costed:

• ~20 hours to do the disparity initialization
• ~20 hours to do the first pass of the bundle optimization (i just did one pass)

Sorry for didn't optimize the speed. Please make sure your computer is powerful enough and yourself is patient enough to run this code.

Zhang, Guofeng, et al. "Consistent depth maps recovery from a video sequence." Pattern Analysis and Machine Intelligence, IEEE Transactions on 31.6 (2009): 974-988.

*For those who are taking this course, please be aware that this repo is just for your reference and plagiarism is a serious offense in nus.