A5083422622- Robotics and Sensor-Based Localization
- 3D Surveying and Cultural Heritage
- Remote Sensing and LiDAR Applications
- Advanced Neural Network Applications
- 3D Shape Modeling and Analysis
- Advanced Vision and Imaging
- Domain Adaptation and Few-Shot Learning
- Optical measurement and interference techniques
- Indoor and Outdoor Localization Technologies
- Human Pose and Action Recognition
- Industrial Vision Systems and Defect Detection
- Computer Graphics and Visualization Techniques
- Video Surveillance and Tracking Methods
- Image Processing and 3D Reconstruction
- Advanced Image and Video Retrieval Techniques
- Robot Manipulation and Learning
Nanyang Technological University
2024-2025
Xiamen University
2019-2023
A point cloud is a set of points defined in 3D metric space. Point clouds have become one the most significant data formats for representation and are gaining increased popularity as result availability acquisition devices, well seeing application areas such robotics, autonomous driving, augmented virtual reality. Deep learning now powerful tool processing computer vision becoming preferred technique tasks classification, segmentation, detection. While deep techniques mainly applied to with...
While Neural Radiance Fields (NeRFs) have advanced the frontiers of novel view synthesis (NVS) using LiDAR data, they still struggle in dynamic scenes. Due to low frequency and sparsity characteristics point clouds, it is challenging spontaneously learn a consistent scene representation from posed scans. In this paper, we propose STGC-NeRF, NeRF method that combines spatial-temporal geometry consistency enhance reconstruction First, temporal regularization regression time-varying geometries...
Unsupervised domain adaptation (UDA) segmentation aims to leverage labeled source data make accurate predictions on unlabeled target data. The key is the network learn domain-invariant representations. In this work, we propose a prototype-guided multitask adversarial (PMAN) achieve this. First, an intensity-aware (IAS-Net) that leverages private intensity information of substantially facilitate feature learning domain. Second, category-level cross-domain alignment strategy introduced flee...
Unsupervised domain adaptation (UDA) is a low-cost way to deal with the lack of annotations in new domain. For outdoor point clouds urban transportation scenes, mismatch sampling patterns and transferability difference between classes make cross-domain segmentation extremely difficult. To overcome these challenges, we propose category-level adversarial framework. Firstly, multi-scale conditioned block that facilitates extract critical low-level domain-dependent knowledge reduce gap caused by...
Despite the significant advancements in pre-training methods for point cloud understanding, directly capturing intricate shape information from irregular clouds without reliance on external data remains a formidable challenge. To address this problem, we propose GPSFormer, an innovative Global Perception and Local Structure Fitting-based Transformer, which learns detailed with remarkable precision. The core of GPSFormer is Module (GPM) Fitting Convolution (LSFConv). Specifically, GPM...
LiDAR-based absolute pose regression estimates the global through a deep network in an end-to-end manner, achieving impressive results learning-based localization. However, accuracy of existing methods still has room to improve due difficulty effectively encoding scene geometry and unsatisfactory quality data. In this work, we propose novel LiDAR localization frame-work, SGLoc, which decouples estimation point cloud correspondence via correspondence. This decoupling encodes because decoupled...
Different domain image sensors or imaging mechanisms provide cross-domain images when sensing the same scene. There is a shift between so that gap different domains major challenge for measuring similarity of feature descriptors extracted from images. Specifically, matching ground camera and unmanned aerial vehicle (UAV) 3-D model-rendered images, which are two kinds extremely challenging way to establish indirectly spatial relationship 2-D spaces. This provides solution virtual-real...
LiDAR localization is of great importance to autonomous vehicles and robotics. Absolute pose regression, directly estimating the mapping from a scene 6-DoF pose, has achieved impressive results in learning-based localization. Different traditional map-based methods, it does not need pre-built 3D map during inference. However, current regression networks typically suffer ambiguities, especially challenging traffic environments, leading large wrong predictions (e.g., outliers) limited...
Point cloud is point sets defined in 3D metric space. has become one of the most significant data format for representation. Its gaining increased popularity as a result availability acquisition devices, such LiDAR, well application areas robotics, autonomous driving, augmented and virtual reality. Deep learning now powerful tool processing computer vision, becoming preferred technique tasks classification, segmentation, detection. While deep techniques are mainly applied to with structured...
Absolute pose regression has shown great potential in LiDAR localization, which learns to regress 6-DoF poses through deep networks. However, recent methods suffer from scene ambiguities challenging scenarios, leading inaccurate and unstable localization. Inspired by neurobiological localization mechanisms, i.e., the firing mechanism of place cells, head-direction grid cells mammalian brains, we propose a novel framework called NIDALoc achieve more robust accurate results. First, Hebbian...