Optics and Precision Engineering, Volume. 33, Issue 6, 928(2025)
Multi-level feature fusion for camera pose regression
Figures & Tables(19)
Fig. 8. Comparison results of TransBoNet and the proposed model on the FULL subset， with a focus on cases A and B.
Fig. 9. Experimental trajectory plots of different algorithmic models on the Oxford RobotCar dataset
Table 1. Comparison of feature encoder structure parameters
View tableView in ArticleTable 1. Comparison of feature encoder structure parameters
卷积阶段 输出尺寸 ResNet101 视觉编码器 卷积层1 卷积层2 卷积层3 卷积层4 卷积层5
Table 2. Training and test sets in Oxford RobotCar
View tableView in ArticleTable 2. Training and test sets in Oxford RobotCar
Sequence Time Tag Mode - 2014-06-26-08-53-56 Overcast Training - 2014-06-26-09-24-58 Overcast Training LOOP1 2014-06-23-15-41-25 Sunny Testing LOOP2 2014-06-23-15-36-04 Sunny Testing - 2014-11-28-12-07-13 Overcast Training - 2014-12-02-15-30-08 Overcast Training FULL1 2014-12-09-13-21-02 Overcast Testing FULL2 2014-12-12-10-45-15 Overcast Testing
Table 3. Experimental results of different algorithms on the Oxford RobotCar dataset
View tableView in ArticleTable 3. Experimental results of different algorithms on the Oxford RobotCar dataset
Method LOOP1 LOOP2 FULL1 FULL2 Average Mean Median Mean Median Mean Median Mean Median Mean Median ResGraphLoc 7.18 3.92 7.08 3.32 16.96 7.43 36.24 6.78 16.87 5.36 PoseNet 25.29 6.88 28.81 5.80 125.6 107.6 131.06 101.8 77.70 55.52 MapNet 8.76 5.79 9.84 4.91 41.4 17.94 59.3 20.04 29.8 12.17 LsG 9.07 - 9.19 - 31.65 - 53.45 - 25.84 - AtLoc 8.61 5.68 8.86 5.05 29.6 11.1 48.2 12.2 23.8 8.54 GNNMapNet 7.76 - 8.15 - 17.35 - 37.81 - 17.77 - AtLoc+ 7.82 4.34 7.24 3.78 21.0 6.40 42.6 7.00 19.7 5.38 TransBoNet - - - - 30.03 6.46 52.42 8.77 41.23 7.62
Table 4. Experimental results of different algorithms on the Oxford RobotCar dataset
View tableView in ArticleTable 4. Experimental results of different algorithms on the Oxford RobotCar dataset
Method LOOP1 LOOP2 FULL1 FULL2 Average Mean Median Mean Median Mean Median Mean Median Mean Median ResGraphLoc 2.48 1.95 2.74 1.86 3.16 1.26 6.52 1.32 3.73 1.60 PoseNet 17.45 2.06 19.62 2.05 27.1 22.5 26.05 20.1 22.56 11.7 MapNet 3.46 1.54 3.96 1.67 12.5 6.68 14.8 6.39 8.68 4.07 LsG 3.31 - 3.53 - 4.51 - 8.60 - 4.99 - AtLoc 4.58 2.23 4.67 2.01 12.4 5.28 11.1 4.63 8.19 3.54 GNNMapNet 2.54 - 2.57 - 3.47 - 7.55 - 4.03 - AtLoc+ 3.62 1.92 3.60 2.04 6.15 1.50 9.95 1.48 5.83 1.74 TransBoNet - - - - 8.18 1.25 10.98 1.69 9.58 1.47
Table 5. Median position/orientation estimation errors on the 7-Scenes dataset
View tableView in ArticleTable 5. Median position/orientation estimation errors on the 7-Scenes dataset
Method Scenes Chess Fire Heads Office Pumpkin Kitchen Stairs Avg GPoseNet 0.20 m，7.11° 0.38 m， 12.3° 0.21 m， 13.8° 0.28 m， 8.83° 0.37 m， 6.94° 0.35 m， 8.15° 0.37 m， 12.5° 0.31 m， 9.95° AtLoc 0.10 m，4.07° 0.25 m， 11.4° 0.16 m， 11.8° 0.17 m， 5.34° 0.21 m， 4.37° 0.23 m， 5.42° 0.26 m， 10.5° 0.20 m， 7.56° AtLoc+ 0.10 m，3.18° 0.26 m， 10.8° 0.14 m， 11.4°0.17 m， 5.16° 0.20 m， 3.94° 0.16 m， 4.90° 0.29 m， 10.2° 0.19 m， 7.08° GNNMapNet 0.08 m，2.82 °0.26 m， 8.94 °0.17 m， 11.41° 0.18 m， 5.08° 0.15 m， 2.77°0.25 m， 4.48° 0.23 m， 8.78° 0.19 m， 6.33° MS-Transformer 0.11 m，4.66° 0.24 m， 9.6°0.14 m， 12.1°0.17 m， 5.66° 0.18 m， 4.44° 0.17 m， 5.94° 0.26 m， 8.45° 0.18 m， 7.28° IRPNet 0.13 m，5.64° 0.25 m， 9.67° 0.15 m， 13.1° 0.24 m， 6.33° 0.22 m， 5.78° 0.30 m， 7.29° 0.34 m， 11.6° 0.23 m， 8.49° DA-model 0.11 m，6.06° 0.26 m， 11.6° 0.15 m， 12.8° 0.17 m， 8.34° 0.21 m， 6.37° 0.19 m， 8.90° 0.26 m， 12.2° 0.19 m， 9.48° TransBoNet 0.11 m，4.48° 0.25 m， 12.4° 0.18 m， 14.0° 0.20 m， 5.08° 0.19 m， 4.77° 0.17 m， 5.35° 0.30 m， 13.0° 0.20 m， 8.45° ResGraphLoc 0.08 m，3.25°0.24 m，9.2°0.15 m， 11.2 °0.15 m，4.55 °0.17 m， 2.63 °0.14 m，3.20 °0.21 m，8.34 °0.16 m，6.05 °
Table 6. Median position/orientation estimation errors on the 4Seasons dataset
View tableView in ArticleTable 6. Median position/orientation estimation errors on the 4Seasons dataset
Method Business Campus Neighborhood Old Town Mean Median Mean Median Mean Median ResGraphLoc 5.25 m，2.20 °2.48 m， 1.45°1.40 m，0.76 °1.22 m，0.65 °25.48 m，2.73 °6.76 m，1.27 °MapNet 10.35 m， 3.78° 5.66 m， 1.83° 2.81 m， 1.05° 1.89 m， 0.92° 46.56 m， 7.14° 16.52 m， 2.12° AtLoc 11.53 m， 4.84° 5.81 m， 1.50° 2.80 m， 1.16° 1.83 m， 0.93° 84.17 m， 7.81° 17.10 m， 1.73° AtLoc+ 13.70 m， 6.41° 5.58 m， 1.94° 2.33 m， 1.39° 1.61 m， 0.88° 68.40 m， 5.51° 14.52 m， 1.69° GNNMapNet 7.69 m， 4.34° 5.52 m， 2.16° 3.02 m， 2.92° 2.14 m， 1.45° 41.54 m， 7.30° 19.23 m， 3.26° IRPNet 10.95 m， 5.38° 5.91 m， 1.82° 3.17 m， 2.85° 1.98 m， 0.90° 55.86 m， 6.97° 17.33 m， 3.11° CoordiNet 11.52 m， 3.44° 6.44 m， 1.38 °1.72 m， 0.86° 1.37 m， 0.69° 43.68 m， 3.58° 11.83 m， 1.36°
Table 7. Model computational complexity analysis
View tableView in ArticleTable 7. Model computational complexity analysis
Method FLOPS/G Activations/millions Params/millions Memory/MB MapNet 3.672 3.7 22.348 89.5 AtLoc 3.068 3.8 24.448 97.9 MS-Transformer 5.578 52.2 18.542 74.5 DA-model 0.237 4.4 3.847 14.2 ReResGraphLoc 0.638 3.6 4.015 21.7
Table 8. Result of ablation experiment
View tableView in ArticleTable 8. Result of ablation experiment
网络配置 误差均值 基础模型 8.42 m，3.68° 基础模型+CoordiBo 7.72 m，2.84° 基础模型+图注意力层 7.68 m，2.72° 基础模型+多级特征融合 7.84 m，3.05° 基础模型+CoordiBo+图注意力层 7.33 m，2.52° 基础模型+CoordiBo+多级特征融合 7.65 m，2.69° ResGraphLoc 7.18 m，2.48°
Tools
Get Citation
Copy Citation Text
Junwen SI, Ziwei ZHOU. Multi-level feature fusion for camera pose regression[J]. Optics and Precision Engineering, 2025, 33(6): 928
Paper Information
Category:
Received: Jul. 13, 2024
Accepted: --
Published Online: Jun. 16, 2025
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20