Related references
Note: Only part of the references are listed.
Article
Computer Science, Information Systems
Xiaoming Zhao et al.
Summary: Existing methods for keypoint detection are not directly optimized using back-propagation due to their non-differentiability. To tackle this issue, we propose a partially differentiable keypoint detection module that accurately outputs sub-pixel keypoints. We introduce a reprojection loss to directly optimize these sub-pixel keypoints and a dispersity peak loss for accurate regularization. Additionally, we employ stable neural reprojection error loss to train sub-pixel descriptors. Our lightweight network achieves high efficiency, running at 95 frames per second on a commercial GPU for 640 x 480 images. The proposed method achieves comparable performance with state-of-the-art approaches in homography estimation, camera pose estimation, and visual (re-)localization tasks while significantly reducing inference time.
IEEE TRANSACTIONS ON MULTIMEDIA
(2023)
Article
Computer Science, Information Systems
Zhihao Shi et al.
Summary: Video frame interpolation aims to synthesize intermediate frames while maintaining spatial and temporal consistencies. Existing methods can be categorized into flow-based and kernel-based. Flow-based methods suffer from inaccuracies in flow map estimation, while kernel-based methods are constrained by rigid kernel shapes. To address these limitations, a novel mechanism called generalized deformable convolution is proposed, which enables data-driven motion learning and flexible sampling in space-time. Experimental results demonstrate the superiority of the new method, particularly for complex motions.
IEEE TRANSACTIONS ON MULTIMEDIA
(2022)
Article
Engineering, Electrical & Electronic
Xiaoming Zhao et al.
Summary: Keypoints matching is a critical component in many image-related applications. This paper proposes integrating the spatial distribution prior of keypoints by utilizing motion estimation from inertial measurement unit (IMU) integration. A probabilistic attention formulation is introduced to incorporate the spatial distribution prior into the attentional graph neural network. The use of spatial distribution prior reduces the network's effort in modeling hidden features, and a projection loss is presented to smooth the edge between matching and un-matching keypoints.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY
(2022)
Article
Robotics
Haosong Yue et al.
Summary: Loop closure detection is a crucial technique to correct accumulated inconsistencies in long-term localization and mapping. We propose a method to build the optimal vocabulary automatically and improve the precision of loop closure detection through a novel topological graph verification method.
JOURNAL OF FIELD ROBOTICS
(2022)
Proceedings Paper
Engineering, Industrial
Yilun Xu et al.
Summary: A novel intensity mapping algorithm, utilizing weighted histogram averaging (WHA) and leveraging the correspondence between histogram bins, is proposed in this paper. The proposed WHA algorithm significantly outperforms the related state-of-the-art intensity mapping methods, addressing the challenges of aligning brightness distribution in images with different exposures.
2022 IEEE 17TH CONFERENCE ON INDUSTRIAL ELECTRONICS AND APPLICATIONS (ICIEA)
(2022)
Proceedings Paper
Computer Science, Artificial Intelligence
Jerome Revaud et al.
Summary: This paper proposes an approach for unsupervised learning of local image descriptors. The approach integrates two matching priors into a single loss function to address the issue of per-pixel correspondence-level supervision required by existing methods. Experimental results show that the proposed approach achieves state-of-the-art performance in various geometric tasks.
2022 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2022)
(2022)
Article
Engineering, Electrical & Electronic
Yiming Ding et al.
Summary: This article proposes a multisource fusion indoor pedestrian autonomous localization framework that combines landmark maps and occupancy grid maps. The short- and medium-term localization accuracy is maintained with the OGI-SLAM algorithm, while the long-term localization accuracy is improved with keypoint maps. Simulation and field trial evaluations show that the system achieves a maximum positioning accuracy of 1.64 m, representing a 57% improvement compared to OGI-SLAM.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2022)
Article
Engineering, Electrical & Electronic
Yuan Rao et al.
Summary: This paper proposes a learning-based hierarchical feature descriptor method that improves the performance of feature matching tasks by leveraging rich complementary clues in intermediate multilevel features. By introducing hierarchical view consistency (HVC) and multiviewer neural network (MVNet), the method effectively utilizes multilevel features and generates discriminative descriptions efficiently.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2022)
Article
Computer Science, Artificial Intelligence
Jiayi Ma et al.
Summary: Image matching is a fundamental task in various visual applications, and with the development of deep learning techniques, there has been an increasing number of methods proposed in this field. However, the challenge remains in choosing the suitable method for specific applications and designing image matching methods with superior performance. This comprehensive review and analysis provide insights into classical and latest techniques, and offer prospects for future development in image matching technologies.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2021)
Proceedings Paper
Computer Science, Artificial Intelligence
Hugo Germain et al.
Summary: The study introduces Neural Reprojection Error (NRE) as a substitute for Reprojection Errors (RE) in absolute camera pose estimation, merging it with feature learning to improve accuracy and robustness.
2021 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, CVPR 2021
(2021)
Proceedings Paper
Computer Science, Artificial Intelligence
Zhen Liu et al.
Summary: This paper introduces a novel attention-guided deformable convolutional network, ADNet, which effectively tackles the challenges of saturation, noise, and misalignments in hand-held multi-frame HDR imaging. By incorporating attention and alignment modules, ADNet achieves state-of-the-art performance in the NTIRE 2021 Multi-Frame HDR Challenge.
2021 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION WORKSHOPS, CVPRW 2021
(2021)
Proceedings Paper
Computer Science, Artificial Intelligence
Suwichaya Suwanwimolkul et al.
Summary: The study introduces a supervised learning approach for keypoint detection using low-level features, achieving state-of-the-art results in 3D reconstruction, especially in improving reprojection error, and exhibiting the highest accuracy in keypoint detection and matching on benchmark datasets through joint learning of descriptors and detectors. A dedicated study on evaluation metrics for keypoint detection and matching is also presented.
2021 IEEE WINTER CONFERENCE ON APPLICATIONS OF COMPUTER VISION WACV 2021
(2021)
Article
Engineering, Electrical & Electronic
Hu Zhang et al.
Summary: The proposed RPI-SURF descriptor integrates local intensity character (LIC) and relative position information (RPI), resulting in significantly improved recall and F-1 score compared to the SURF descriptor. In real froth image pairs, the RPI-SURF descriptor detected accurate matches increased by at least nearly 60%.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2021)
Article
Computer Science, Software Engineering
Arno Knapitsch et al.
ACM TRANSACTIONS ON GRAPHICS
(2017)
Proceedings Paper
Computer Science, Artificial Intelligence
Jifeng Dai et al.
2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV)
(2017)
Proceedings Paper
Computer Science, Artificial Intelligence
Vassileios Balntas et al.
30TH IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2017)
(2017)
