研究目的
To learn a similarity metric of patches from reference and target images for 3-D image registration, such that patches with small projection error receive high similarity scores, improving pose optimization convergence.
研究成果
The proposed networks (classification, regression, rank) achieve superior performance in learning a similarity metric for patch-based image registration compared to traditional methods like SSD and SAD. This can improve convergence speed and robustness in visual geometry applications such as visual odometry. Future work includes applying networks to intensity-based registration and improving data set quality.
研究不足
The networks are trained on specific data sets (TUM), and the accuracy is hindered by errors in data sets. The patch size is fixed at 32x32, and the method relies on the quality of collected data. Future improvements could involve higher precision data sets or virtual data.
1：Experimental Design and Method Selection:
Designed and trained classification, regression, and rank neural networks to learn a similarity metric for patch-based image registration. Used self-collected data sets from RGB-D and monocular images.
2：Sample Selection and Data Sources:
Collected patch data sets from TUM RGB-D data sets and TUM monocular data sets using methods based on RGB-D, DSO, and improved DSO. Patches are 32x32 in size, extracted around fast corners or pattern centers.
3：List of Experimental Equipment and Materials:
Used TUM RGB-D data sets, TUM monocular data sets, and DSO system for data collection. No specific equipment models or brands mentioned.
4：Experimental Procedures and Operational Workflow:
Detected fast corners, extracted patches, applied transformations with interference, computed projection errors as labels, preprocessed data by discarding patches with high SSD, trained networks with specified loss functions, and evaluated performance using designed criteria.
5：Data Analysis Methods:
Used evaluation criteria like VDCG and pVDCG, standard error charts, and comparison with traditional methods (SSD, SAD).
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
