Fig. 1. Depth-dependent radius of blur kernel
Fig. 2. Data samples in unbalanced defocus stereo vision dataset. (a1)-(a6) Left and right blur images, left and right clear images, and left and right disparity maps of No. 0 data; (b1)-(b6) left and right blur images, left and right clear images, and left and right disparity maps of No. 2000 data
Fig. 3. Visualized deblurred results of synthetic image. (a1)-(a5) Left blur image, deblurred left images by Nah, DavaNet, and BLNet, and clear left image of No. 0 data; (b1)-(b5) right blur image, deblurred right images by Nah, DavaNet, and BLNet, and clear right image of No. 0 data; (c1)-(c5) left blur image, deblurred left images by Nah, DavaNet, and BLNet, and clear left image of No. 2000 data; (d1)-(d5) right blur image, deblurred right images by Nah, DavaNet, and BLNet, and clear right image of No. 2000 data; (e1)-(e5) left blur image, deblurred left images by Nah, DavaNet, and BLNet, and clear left image of No. 4000 data; (f1)-(f5) right blur image, deblurred right images by Nah, DavaNet, and BLNet, and clear right image of No. 4000 data
Fig. 4. Visualized deblurred results of real-scene images in Middlebury 2014 dataset. (a1)-(a5) Left blur image, deblurred left images by Nah, DavaNet, and BLNet, and clear left image of Adirondack; (b1)-(b5) right blur image, deblurred right images by Nah, DavaNet, and BLNet, and clear right image of Adirondack; (c1)-(c5) left blur image, deblurred left images by Nah, DavaNet, and BLNet, and clear left image of Motorcycle; (d1)-(d5) right blur image, deblurred right images by Nah, DavaNet, and BLNet, and clear right image of Motorcycle
Fig. 5. Visualized stereo matching results of synthetic image. (a1)-(a5) Left and right blur images, disparity maps of PSMNet-C and PSMNet-B, and ground-truth disparity map of No. 0 data; (b1)-(b5) left and right blur images, disparity maps of PSMNet-C and PSMNet-B, and ground-truth disparity map of No. 2000 data; (c1)-(c5) left and right blur images, disparity maps of PSMNet-C and PSMNet-B, and ground-truth disparity map of No. 4000 data
Fig. 6. Visualized stereo matching results of real-scene images in Middlebury 2014 dataset. (a1)-(a5) Left and right blur images, disparity maps of PSMNet-C and PSMNet-B, and ground-truth disparity map of Adirondack; (b1)-(b5) left and right blur images, disparity maps of PSMNet-C and PSMNet-B, and ground-truth disparity map of Motorcycle
Fig. 7. Test on real defocus blur images. (a) Stereo vision cameras; (b) experimental scene for test; (c)(d) left and right blur images; (e)(f) deblurred left and right images by BLNet; (g) disparity map calculated by PSMNet-B; (h) reconstructed 3D point clouds
Number | Nah | DavaNet | BLNet |
---|
PSNR | SSIM | PSNR | SSIM | PSNR | SSIM |
---|
Average | 35.51 | 0.95 | 34.66 | 0.93 | 32.75 | 0.93 | 0 | 36.30 | 0.94 | 35.50 | 0.93 | 33.49 | 0.93 | 500 | 39.20 | 0.97 | 37.87 | 0.96 | 35.38 | 0.96 | 1000 | 37.22 | 0.97 | 36.74 | 0.96 | 33.58 | 0.95 | 1500 | 34.47 | 0.93 | 33.42 | 0.92 | 32.29 | 0.91 | 2000 | 36.02 | 0.95 | 34.97 | 0.94 | 32.44 | 0.93 | 2500 | 35.00 | 0.94 | 33.99 | 0.92 | 31.46 | 0.91 | 3000 | 33.54 | 0.92 | 32.80 | 0.90 | 31.95 | 0.90 | 3500 | 33.52 | 0.93 | 32.71 | 0.92 | 31.65 | 0.92 | 4000 | 34.32 | 0.96 | 33.93 | 0.95 | 32.47 | 0.94 |
|
Table 1. Deblurred results of synthetic image
Scene | Nah | DavaNet | BLNet |
---|
PSNR | SSIM | PSNR | SSIM | PSNR | SSIM |
---|
Average | 33.97 | 0.93 | 32.77 | 0.91 | 30.30 | 0.90 | Adirondack | 38.28 | 0.96 | 36.27 | 0.94 | 32.68 | 0.94 | Motorcycle | 30.01 | 0.92 | 29.55 | 0.91 | 27.47 | 0.90 | Piano | 35.36 | 0.95 | 33.82 | 0.92 | 31.93 | 0.92 | Pipes | 31.79 | 0.90 | 31.55 | 0.89 | 30.00 | 0.89 | Playroom | 32.00 | 0.94 | 31.04 | 0.91 | 30.55 | 0.91 | Playtable | 33.31 | 0.87 | 30.97 | 0.83 | 29.35 | 0.82 | Recycle | 37.54 | 0.96 | 36.10 | 0.94 | 31.00 | 0.95 | Shelves | 33.45 | 0.94 | 32.88 | 0.92 | 29.41 | 0.92 |
|
Table 2. Deblurred results of real-scene images in Middlebury 2014 dataset
Number | PSMNet-C | PSMNet-B |
---|
D3 /% | EPE /pixel | D3 /% | EPE /pixel |
---|
Average | 34.49 | 6.96 | 9.16 | 2.01 | 0 | 40.90 | 9.25 | 13.20 | 2.63 | 500 | 21.30 | 3.16 | 2.20 | 0.74 | 1000 | 14.90 | 5.63 | 10.30 | 3.62 | 1500 | 42.60 | 9.71 | 13.50 | 2.53 | 2000 | 46.50 | 11.61 | 10.20 | 1.71 | 2500 | 29.30 | 5.53 | 7.20 | 1.39 | 3000 | 37.90 | 4.82 | 13.10 | 2.73 | 3500 | 37.30 | 6.77 | 8.30 | 1.81 | 4000 | 39.70 | 6.20 | 4.40 | 0.90 |
|
Table 3. Stereo matching results of the synthetic data
Scene | PSMNet-C | PSMNet-B |
---|
D3 /% | EPE /pixel | D3 /% | EPE /pixel |
---|
Average | 43.73 | 7.21 | 26.48 | 5.64 | Adirondack | 42.50 | 7.11 | 15.20 | 2.97 | Motorcycle | 40.10 | 7.06 | 20.70 | 4.42 | Piano | 33.40 | 3.89 | 27.20 | 5.45 | Pipes | 57.20 | 13.56 | 25.60 | 6.95 | Playroom | 44.60 | 7.92 | 39.30 | 9.91 | Playtable | 37.00 | 5.16 | 25.30 | 5.87 | Recycle | 31.90 | 3.96 | 17.80 | 2.77 | Shelves | 63.10 | 9.01 | 40.70 | 6.81 |
|
Table 4. Stereo matching results of real-scene images from Middlebury 2014 dataset