A5051282546- Advanced Vision and Imaging
- Advanced Image Processing Techniques
- Human Pose and Action Recognition
- Image Processing Techniques and Applications
- Robotics and Sensor-Based Localization
- Image Enhancement Techniques
- Advanced Image and Video Retrieval Techniques
- 3D Shape Modeling and Analysis
- Video Surveillance and Tracking Methods
- Graph Theory and Algorithms
- Advanced Graph Neural Networks
- Optical measurement and interference techniques
- Advanced Clustering Algorithms Research
Harbin Engineering University
2021-2025
Ministry of Industry and Information Technology
2022-2024
Recurrently refining the optical flow based on a single high-resolution feature demonstrates high performance. We exploit strength of this strategy to build novel architecture for joint learning and depth. Our pro-posed is improved work in case training unlabeled data, which extremely challenging. The loss computed iterations carried out over feature, where reconstruction fails optimize accuracy particularity occluded regions. Therefore, we propose hierarchically refine across multiple...
Scene flow describes the 3D motion in a scene. It can be modeled as single task or composite of auxiliary tasks depth, camera motion, and optical estimation. Deep learning's emergence recent years has broadened horizons for new methodologies estimating these tasks, either separate joint to reconstruct scene flow. The sequence images that are synthesized captured by is used input methods, which face challenge dealing with various situations provide most accurate such image quality. Nowadays,...
In this work, we propose a self-supervised scene flow framework for joint learning of optical flow, stereo depth, camera pose, and rigidity map handle the occlusion during training. Specifically, feature masking method to alleviate impact on correlation result reduce outliers in both depth map. We use improved estimate motion directly using Perspective-n-Point method, which improves it accordingly. Furthermore, recursively update occluded non-occluded regions with cues learned from rigid...
Optical flow is the process of estimating motion in scenes. Each object scene has a homogeneous motion, i.e., moves same direction with velocity. Therefore, connecting parts an image globally provides essential cue for learning accurate motion. Convolution-based methods estimate features from local regions, which miss this important cue. Recently, some used Transformer to model global dependencies improve optical flow. However, suffers excessive attention computations and still brings...
Scene flow estimation is the task of obtaining 3D motion from a dynamic scene. Due to sparseness point clouds, extracting features for local group points separately may result in different that all belong same object. This difference makes global correlation prone producing an unacceptable flow. Local restricts algorithm capturing limited movements and fails when fast movement or large deformation object occurs. Therefore, we propose transformer-based scene method can perform feature...