dfusmc's Introduction

DfUSMC

(DfUSMC: Depth from Uncalibrated Small Motion Clip)

Source code and datasets for the paper:

H. Ha, S. Im, J. Park, H.-G. Jeon and I.S. Kweon, High-quality Depth from Uncalibrated Small Motion Clip, CVPR 2016

Project page

Dependencies

How to Run

Put your small motion clip in "Dataset" folder
Run (with sudo)

cmake .
make
./DfUSMC data_name video_extension
(for example, ./DfUSMC Bikes avi)

Important Information

Frame selection

The current implementation uses only the first 30 frames of your video clip. If you want to try with a different number of images or different sampling rate, please modify the "LoadSmallMotionClip" function.

Dense matching step

We implemented the function "DenseMatching" to receive a scale for image downsampling and the number of labels for your convenience in testing. (Default: 0.5 scale and 64 labels for quick tests, but please remind that 1.0 and 256 were used in the paper)

For the depth refinement, we utilized a tree-based depth upsampling approach [1,2].

Authors

IMPORTANT: If you use this software please cite the following in any resulting publication:

@inproceedings{ha2016high,
  title={High-quality Depth from Uncalibrated Small Motion Clip},
  author={Ha, Hyowon and Im, Sunghoon and Park, Jaesik and Jeon, Hae-Gon and Kweon, In So},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2016}
}

References

Yang, Qingxiong. "Stereo matching using tree filtering." IEEE transactions on pattern analysis and machine intelligence 37.4 (2015): 834-846.
Yang, Qingxiong. "A non-local cost aggregation method for stereo matching." Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on. IEEE, 2012.

dfusmc's Issues

question

for iphone6 ，the focal length of taking pictures and videos is not the same，how do youget the true focal length？

d->u and u->d mapping

Hi~
I'm getting little confused with the distorted->undistorted mapping:
remap(img_dist, img_undist, ud_mapx, ud_mapy, CV_INTER_CUBIC);
Why is ud_mapx/y used to apply d->u mapping?
Could anybody pls tell me? Thanks.

Could you please share the entire dataset?

I found some datasets and ground truth depth in the paper are not in the "dataset" folder. Could you please upload them? Thanks!

questions for depth？

hi ,i run you code and test my own data(iphone6 video fixed paramters ),many question:
1.for self-calibrate.
our self-calibrate result : f=2006.5 but your result is 1278,,,,.how do you get the phone real F value while iphone record the video .
and then,i using Radial distortion and tangential distortion model (k1,k2,k3,p1,p2).but the result is Fx=2084.4 Fy=2067.3. not the 1278.
2.for result
the result refine_xxx.bmp is disparity d not the depth z? z=F*Baseline/d?