基于改进代价计算和视差候选策略的立体匹配 
下载： 978次
Stereo Matching Based on Improved Cost Calculation and a Disparity Candidate Strategy
图 & 表
图 1. 算法流程
Fig. 1. Algorithm flow
下载图片 查看原文
图 2. L-Census编码过程
Fig. 2. Coding process of L-Census
下载图片 查看原文
图 3. 物体边界区的失效示意图(p的正确匹配点是q,通过(3)式计算得到的误匹配点为e)。(a)Teddy左局部图; (b)Teddy右局部图
Fig. 3. Failure diagram in the boundary area of the object, in which the correct matching point of p is q, but the corresponding mismatching point is e obtained by Eq. (3). (a) Part of left image of Teddy; (b) part of right image of Teddy
下载图片 查看原文
图 4. 双向梯度代价自适应加权结合策略有效性图示
Fig. 4. Illustration of the effectiveness of the adaptive weighted combining strategy of bidirectional gradient cost
下载图片 查看原文
图 5. 通过不同的横向和纵向梯度代价组合方法获得的视差图。(a)原图;(b)真实视差图;(c)采用(3)式得到的视差图;(d)ABiGrad得到的视差图
Fig. 5. Disparity maps obtained by different horizontal and vertical gradient costs combining methods. (a) Original image; (b) real disparity map; (c) disparity map obtained using Eq. (3); (d) disparity map obtained using ABiGrad
下载图片 查看原文
图 6. 基于十字交叉的自适应窗口的构建。 (a)十字臂; (b)自适应支持区域
Fig. 6. Construction of adaptive cross window. (a) Cross arm; (b) adaptive support area
下载图片 查看原文
图 7. 验证“候选视差”思路的区域
Fig. 7. Area to verify the "candidate disparities" idea
下载图片 查看原文
图 8. 视差计算流程图
Fig. 8. Flow chart of disparity calculation
下载图片 查看原文
图 9. 视差计算伪代码
Fig. 9. Pseudo code for disparity calculation
下载图片 查看原文
图 10. 参数变化对误匹配率的影响。(a)误匹配率随λGrad变化;(b)误匹配随λCensus变化;(c)误匹配率随M变化;(d)误匹配率随τc变化;(e)误匹配率随τd变化
Fig. 10. Effect of parameters changing on error rate. (a) Variation of error rate with λGrad; (b) variation of error rate with λCensus; (c) variation of error rate with M; (d) variation of error rate with τc; (e) variation of error rate with τd
下载图片 查看原文
图 11. 不同代价计算方法获得的视差图。 (a)参考图; (b)真实视差图; (c)AD-Cen; (d)AD-Grad; (e)LCen-ABiGrad
Fig. 11. Disparity maps obtained by different cost calculation methods. (a) Reference image; (b) real disparity map; (c) AD-Cen; (d) AD-Grad; (e) LCen-ABiGrad
下载图片 查看原文
图 12. DC相对WTA在重复纹理区(直线框)、弱纹理区(虚线框)和无纹理区(双直线框)的优势。(a)参考图;(b)WTA获取的视差图;(c)DC获取的视差图;(d)标记视差图
Fig. 12. Advantages of DC over WTA in the repeated texture area (straight frame), weak texture area (dotted frame) and untextured area (double straight frame). (a) Reference image; (b) disparity map obtained by WTA; (c) disparity map obtained by DC; (d) marked disparity map
下载图片 查看原文
图 13. 本文算法在标准立体图像对上的结果。(a)参考图;(b)真实视差图;(c)本文算法生成的视差图;(d)误匹配图
Fig. 13. Results of our algorithm on standard stereo image pairs. (a) Reference image; (b) real disparity map; (c) disparity map generated by our algorithm; (d) mismatching map
下载图片 查看原文
表 1各候选视差的正确率
Table1. Correct rate of each candidate disparity unit: %
| Candidate disparity | d(1) | d(2) | d(3) | d(4) | Others | Sum |
|---|
| Proportion | 53.3 | 12.1 | 6.0 | 3.7 | 24.9 | 100.0 |
|
查看原文
表 2实验参数设置
Table2. Experimental parameter setting
| Parameter | L1 | L2 | τ1 | τ2 | τ3 | λCensus | λGrad | τVN | τVR | M | τc | τd |
|---|
| Value | 17 | 34 | 20 | 6 | 20 | 13 | 1 | 20 | 0.4 | 2 | 1.09 | 10 |
|
查看原文
表 3不同代价计算方法的误匹配率
Table3. Mismatching rate of different cost calculation methods unit: %
| Algorithm | Tsukuba | Teddy | Art | Moebius | Books | Wood1 | Cloth2 | Laundry | |
|---|
| AD-Cen | 4.48 | 15.20 | 31.80 | 20.80 | 24.40 | 26.20 | 18.20 | 32.80 | | | AD-Grad | 4.35 | 17.70 | 32.00 | 22.00 | 25.00 | 26.60 | 18.80 | 30.30 | | | LCen-AbiGrad | 4.06 | 15.10 | 30.50 | 18.20 | 21.50 | 24.80 | 18.00 | 27.90 | | | Algorithm | Bowling1 | Baby1 | Aloe | Lampshade1 | Midd1 | Rocks1 | Wood2 | Reindeer | Ave(all) | | AD-Cen | 31.90 | 15.00 | 16.50 | 23.40 | 42.60 | 13.90 | 15.60 | 30.10 | 22.70 | | AD-Grad | 35.80 | 16.60 | 18.90 | 23.60 | 43.30 | 13.30 | 15.20 | 30.20 | 23.40 | | LCen-ABiGrad | 26.00 | 15.00 | 17.10 | 20.00 | 24.30 | 13.00 | 14.50 | 26.70 | 19.80 |
|
查看原文
表 4不同视差计算策略的误匹配率
Table4. Mismatching rate of different disparity calculation strategies unit: %
| Algorithm | Teddy | Dolls | Reindeer | Baby2 | Bowling2 | Cloth2 | Aloe |
|---|
| WTA | 15.10 | 18.00 | 26.70 | 17.20 | 24.00 | 18.00 | 17.10 | | SO | 14.40 | 17.50 | 24.80 | 20.00 | 24.00 | 17.80 | 15.00 | | DC | 14.80 | 17.90 | 26.20 | 16.70 | 23.70 | 17.50 | 16.90 | | Algorithm | Flowerpots | Midd1 | Midd2 | Plastic | Rocks2 | Rocks1 | Ave(all) | | WTA | 23.70 | 24.30 | 23.90 | 34.90 | 13.20 | 13.00 | 20.70 | | SO | 25.60 | 21.50 | 19.10 | 36.70 | 12.90 | 13.00 | 20.18 | | DC | 23.20 | 23.10 | 23.00 | 34.90 | 12.90 | 12.80 | 20.28 |
|
查看原文
表 5不同算法在标准立体图片对上的误匹配率
Table5. Mismatch rate of different algorithms on standard stereo picture pairs unit: %
| Algorithm | Tsukuba | Venus | Teddy | Cones | Averageerror |
|---|
| N-occ | All | Disc | N-occ | All | Disc | N-occ | All | Disc | N-occ | All | Disc |
|---|
| Ours | 2.12 | 2.50 | 8.23 | 0.25 | 0.62 | 2.24 | 4.97 | 11.00 | 12.50 | 2.78 | 8.68 | 8.04 | 5.33 | | SO+borders | 1.29 | 1.71 | 6.83 | 0.25 | 0.53 | 2.26 | 7.02 | 12.20 | 16.30 | 3.90 | 9.85 | 10.20 | 6.03 | | Assw-Grad | 1.57 | 2.00 | 7.32 | 0.89 | 1.00 | 3.18 | 7.20 | 12.40 | 16.10 | 3.68 | 9.18 | 8.62 | 6.10 | | GradAdaptWgt | 2.26 | 2.63 | 8.99 | 0.99 | 1.39 | 4.92 | 8.00 | 13.10 | 18.60 | 2.61 | 7.67 | 7.43 | 6.55 | | AdaptAggrDP | 1.57 | 3.50 | 8.27 | 1.53 | 2.69 | 12.40 | 6.79 | 14.30 | 16.20 | 5.53 | 13.20 | 14.80 | 8.40 |
|
查看原文
表 6不同算法在标准立体图片对上的运行时间
Table6. Running time of different algorithms on standard stereo image pairs unit: s
| Algorithm | Tsukuba | Venus | Teddy | Cones |
|---|
| Ours | 0.9 | 1.4 | 3.5 | 3.3 | | Assw-Grad | 1.7 | 2.8 | 4.2 | 3.9 | | GradAdaptWgt | 24 | 39 | 59 | 59 |
|
查看原文
宋巍, 魏新宇, 张明华, 贺琪. 基于改进代价计算和视差候选策略的立体匹配[J]. 激光与光电子学进展, 2021, 58(2): 0215001. Wei Song, Xinyu Wei, Minghua Zhang, Qi He. Stereo Matching Based on Improved Cost Calculation and a Disparity Candidate Strategy[J]. Laser & Optoelectronics Progress, 2021, 58(2): 0215001.
PDF全文
