Author Affiliations
School of Automation, Harbin University of Science and Technology, Harbin, Heilongjiang 150080, Chinashow less
Fig. 1. Image to be matched taken by mobile phone
Fig. 2. Matching results of different algorithms. (a) Harris algorithm; (b) SIFT algorithm; (c) Harris-SIFT algorithm; (d) our algorithm
Fig. 3. Experimental images at different l. (a) l=3; (b) l=2; (c) l=1
Fig. 4. Structure of the ResNet50
Fig. 5. Some experimental images
Fig. 6. Sparse three-dimensional point cloud. (a) SFM algorithm; (b) improved SFM algorithm
Fig. 7. Dense point cloud of the object. (a) MVS algorithm; (b) our algorithm
Fig. 8. Three-dimensional reconstruction results of the Poisson surface. (a) Original algorithm; (b) improved algorithm
Matching algorithm | Harris | SIFT | Harris-SIFT | Improved Harris-SIFT |
---|
Left image feature point | 2057 | 2786 | 1490 | 1450 | Right image feature point | 2172 | 2801 | 1612 | 1597 | Matching point | 1600 | 821 | 506 | 492 | Matching accuracy /% | 81.61 | 93.54 | 93.67 | 94.92 | Matching time /s | 5.02 | 11.64 | 3.94 | 3.61 |
|
Table 1. Accuracies and running time of 4 matching algorithms
l | Number of image matches | Match time /s | Number of point clouds |
---|
1 | 32719 | 94 | 101823 | 2 | 24589 | 61 | 99745 | 3 | 15541 | 42 | 65521 |
|
Table 2. Image parameters at different l
Algorithm | XRMSE | YREL | |
---|
MVS | 2.216 | 0.101 | 0.075 | Ours | 1.123 | 0.081 | 0.038 |
|
Table 3. Errors of different algorithms
Algorithm | Time /min | Point cloud | SSIM |
---|
SFM+MVS | 8.21 | 228166 | 0.53 | Improve algorithm | 6.48 | 261617 | 0.81 |
|
Table 4. Three-dimensional reconstruction results of different algorithms