An Efficient Dense Descriptor Applied for Wide Baseline Stereo
We show that it is possible to estimate depth from two wide baseline images using a dense descriptor. Our local descriptor, called DAISY, is very fast and efficient to compute. It depends on histograms of gradients like SIFT and GLOH but uses a Gaussian weighting and circularly symmetrical kernel. This gives us our speed and efficiency for dense computations. We compute 200-length descriptors for every pixel in an 800×600 image in less than 5 seconds.
DAISY has also been used for multiview stereo reconstruction
Herz-Jesu Grid. By using two images, one from the left-most column and one from the upper row, we compute depth and occlusion maps from the view point of the row image. In the diagonal, we display the ground truth depth maps. We marked the correctly detected occlusions with green, incorrectly detected ones with blue and the missed ones with red. From this Figure, it is apparent that DAISY can handle quite large baselines without losing too much from its accuracy.
Stereo reconstructions obtained using the DAISY descriptor. The first row are the input images where we use the first image and one of the other images as input and the second row shows the respective reconstructions where the first one is the laser scanned ground truth image.
you can download the images that we’ve used in the publications from here. archive contains all the images with calibrations and the groundtruth depth maps used in the experiments. please give a reference to the papers if you use the data in the set. additionally, if you’re using the ground truth depth maps, you need to cite mvs site too.
dataset (468 mb) //Friday, May 28, 2010 14:15:51 +0200
Source code is available under the BSD License.
daisy-v1.8.1 // Sunday, October 11, 2009 11:42:22 +0200