Faculty Publications
Permanent URI for this communityhttps://idr.nitk.ac.in/handle/123456789/18736
Publications by NITK Faculty
Browse
2 results
Search Results
Item Online Video Stabilization using Mesh Flow with Minimum Latency(Institute of Electrical and Electronics Engineers Inc., 2023) Devaguptam, D.; Thanmai, K.; Raj, L.S.; Naik, D.; Kolkar, R.Most existing video stabilization techniques are used for post-processing, where previously recorded videos are given to the model to obtain stabilized versions. Online video stabilization usually relies on sensors like gyroscopes or assumes constant motion, which is not suitable for videos with changing motions. This work introduces a video stabilization technique with just one-frame latency. The algorithm operates at the spatial level in the infrequent domain, tracking the motion of mesh vertices. Motion tracks of feature marks are combined with the nearest mesh vertex using two median gauges, assigning each vertex a smooth motion track. The proposed approach, called anticipated foster track leveling, smoothes the motion profiles by utilizing previous motions and adapting accordingly for smoother results. This method can handle changes in movement in space and time and works in real-time, allowing applications in security systems, robotics, and unmanned aerial vehicles (UAVs). When evaluated against other models, MeshFlow gives an overall good performance in all comparison metrics evaluated. Hence MeshFlow can be used as a reliable low-latency technique for real-time video stabilization in remote devices. © 2023 IEEE.Item Node Classification in Weighted Complex Networks Using Neighborhood Feature Similarity(Institute of Electrical and Electronics Engineers Inc., 2024) Shetty, R.D.; Bhattacharjee, S.; Thanmai, K.The potential of graph representation learning schemes has attained great acceptance in diverse, complex network applications. Most of the existing Graph Neural Network (GNN) architectures explore the node features aggregation and feature transformation within the neighborhoods, mainly performed on the unweighted graphs. Also, the existing GNN architectures consider all sets of neighborhood features, which are computationally expensive tasks. Practically, most of the real-world graphs are weighted graphs, and it is important to learn the representation of weighted graphs. In this work, we generate and leverage information of the best possible feature combination from the multiple levels of the networks. Edge weights and the connection structure are considered for generating node embedding, and classifying the node more accurately. The proposed framework, Similarity Feature Embedding GNN (SFEGNN), can be efficiently used for node classification in the weighted networks by leveraging feature overlap similarity from the network structure. This novel approach is helpful in modeling weighted networks for node classification and determining how strongly the neighborhood features are correlated. We validate the efficacy of SFEGNN on six benchmark datasets with varying degrees of homophily ratio and found that it is effective even for highly heterophily networks. Our model has empirically outperformed the state-of-the-art GNN framework with the highest accuracy improvement of 28.88%. © 2017 IEEE.