Download presentation
1
Fast and Accurate Image Matching with Cascade Hashing for 3D Reconstruction
J Cheng et al,. CVPR14 Hyunchul Yang( 양현철 ) Jian Cheng* (Chinese Academy of Sciences), Cong Leng (Chinese Academy of Science), Jiaxiang Wu (Chinese Academy of Sciences), Hainan Cui, Hanqing Lu (NLPR , IACAS) Hello everyone. I’ll start my presentation. The title is fast and accurate image matching with cascade hashing for 3d reconstructions
2
Background Related work Approach Experiments Result Conclusion
Overview Background Related work Approach Experiments Result Conclusion
3
3D Reconstruction technology is similar with image retrieval
Background 3D Reconstruction technology is similar with image retrieval
4
Feature Matching is very computational cost in 3D reconstruction
Background Feature Matching is very computational cost in 3D reconstruction Around 50% Running time Feature Extraction Feature Matching (Hashing) Track Generation Geometric Estimation 3D Reconstruction
5
Related work KD-Tree Well known nearest neighbor search algorithm.
But it not suitable for high dimensional space
6
Related work LDAHash(Linear Discriminant Analysis) PAMI 2012
7
Related work LDAHash(Linear Discriminant Analysis) PAMI 2012
8
Related work LDAHash(Linear Discriminant Analysis) PAMI 2012
9
P is projection matrix that is designed either
Related work LDAHash(Linear Discriminant Analysis) PAMI 2012 Px + t = 0 P is projection matrix that is designed either to solely minimize the in-class covariance of the descriptor or to jointly minimize the in-class covariance and maximize the covariance across classes t is threshold matrix so that the resulting binary strings maximize recognition rates.
10
3. Approach
11
Approach Cascade Hashing ( 3 Step ) Coarse search
Hashing Lookup with Multiple tables Hashing Remapping Top k Ranking via Hashing Coarse search
12
Approach Cascade Hashing ( 3 Step ) Refined search
Hashing Lookup with Multiple tables Hashing Remapping Top k Ranking via Hashing Refined search
13
Approach Cascade Hashing ( 3 Step ) Brute search
Hashing Lookup with Multiple tables Hashing Remapping Top k Ranking via Hashing Brute search
14
Approach Hashing Lookup with Multiple tables 1st hash table
L = Try Count, Number of tables m = Number of hyper-planes
15
Approach Hashing Lookup with Multiple tables 2nd hash table
L = Try Count, Number of tables m = Number of hyper-planes
16
Approach Hashing Lookup with Multiple tables Lth hash table
L = Try Count, Number of tables m = Number of hyper-planes
17
Approach Hashing Lookup with Multiple tables Ex) m = 8 L = 6
L = Number of tables m = Number of hyper-planes Ex) m = 8 L = 6 Coarse search
18
Approach 2. Hashing Remapping
19
Approach 2. Hashing Remapping n = Number of hyper-planes
20
Approach 2. Hashing Remapping n = Number of hyper-planes 3 1 2 2 3 3 2
3 2 2 4
21
Approach 2. Hashing Remapping Ex) n = 128 Refined search
n = Number of hyper-planes Ex) n = 128 Refined search
22
Approach 3. Top k Ranking via Hashing
23
Approach 3. Top k Ranking via Hashing
Brute-force search on top k bucket
24
4. Result
25
Result Standard Oxford dataset with SIFT key points
26
Result Standard Oxford dataset with SIFT key points x 10
27
Result Standard Oxford dataset with SIFT key points x 15
28
Result Standard Oxford dataset with SIFT key points x 100
29
Conclusion Paper proposed a Cascade Hashing method to speed up the image matching Accelerated by our approach in hundreds times than brute force matching Even achieves ten times or more than Kd-tree based matching While retaining comparable accuracy. How about apply this idea to spherical hashing?
30
Thank you Q & A
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.