Fast Matting Using Large Kernel Matting Laplacian Matrices

Supplementary Materials

 

Faster propagation

input trimap r=1 r=17

The propagation speed is much faster when we use a larger kernel. Notice that each iteration takes the same time in our algorithm (independent of r). A large kernel greatly reduces the whole running time due to fewer iterations.

 

KD-tree trimap segmentation

image

segmented trimap

Once the trimap is segmented, we solve the matte in each segment. This figure illustrates the order of solving the segments. The kernel size is adaptively set. (The outcome here is the Local Step 1 in our paper.)

 

High resolution results (Mega-pixel images)

Images and trimaps are from the data set in www.alphamatting.com

(Click the images to see the full size)

7.6M-pixel image (3280*2310)

(This is the high resolution version of Fig. 1 in our paper.)

trimap

closed-form using coarse-to-fine

1359s

ours

48s

7.8M-pixel image (3355*2315)

trimap

closed-form using coarse-to-fine

98s

ours

10.0s

5.4M-pixel image (2090*2600)

trimap

closed-form using coarse-to-fine

273s

ours

15.5s

6.7M-pixel image (3173*2100)

trimap

closed-form using coarse-to-fine

140s

ours

11.1s

7.6M-pixel image (2908*2600)

trimap

closed-form using coarse-to-fine

218s

ours

14.7s