Duxin Zhu

Duxin Zhu

Auto-generated profile, based on publicly available information

Duxin Zhu

School of Computer Engineering, Jimei University, Xiamen, Fujian, China

Research Focus & Metrics

Representative Publications

View All Publications
18 October 2024
Verified
Conference Paper
Open access
JLJinyuan Li
Jinyuan Li
DZDuxin Zhu
Duxin Zhu
ZLZhuangzhi Liu
Zhuangzhi Liu
YLYang Luo
Yang Luo
JSJinhe Su
Jinhe Su
YWYundong Wu
Yundong Wu
3D Visual Grounding (3D VG) has become increasingly important in applications such as robotics and human-computer interaction. However, accurately localizing target objects in complex 3D scenes based on natural language descriptions remains a significant challenge. While current methods based on DETR have shown notable improvements in the field of 3D VG, they struggle to achieve precise localization in the presence of multiple objects. To address these challenges, we propose Progressive Proposal Detection Transformer for 3D VG (PP-DETR). We introduce a progressive proposal generation strategy into the DETR-based 3D VG model which is named Progressive Category-Constrained Sampler(PCCS). After thoroughly integrating point clouds features with natural language features, our method generates object queries through a stepwise sampling of point clouds features. These queries are then decoded to produce the final bounding boxes for the target objects. PCCS reduces the interference from irrelevant proposals, enhancing the model's focus on relevant objects and thereby improving overall localization accuracy. Extensive experiments conducted on the ScanRefer dataset demonstrate that our model achieves an overall accuracy of 54.06% on the overall0.25 metric, surpassing the baseline EDA by 0.23%. In more complex multi-object scenarios, our approach yields improvements of 0.45% and 0.49% in the multi0.25 and multi0.5 metrics, respectively.

18 October 2024
Verified
Conference Paper
Open access
QHQiaoqiao Hao
Qiaoqiao Hao
ZLZhaoliang Liu
Zhaoliang Liu
DZDuxin Zhu
Duxin Zhu
JSJinhe Su
Jinhe Su
ZGZheng Gong
Zheng Gong
GCGuorong Cai
Guorong Cai
Significant progress has been made in roof wireframe reconstruction. However, previous work faced a major issue of positive-negative sample imbalance in edge detection using fully connected graphs of corner points. This led to incomplete roof structures and redundant edges. In this paper, we introduce a new learning point for edge attraction fields, called LEAF, for roof wireframe reconstruction. We observed that points distributed along the edges determine whether there is a connection between diagonal points. To address this, we designed an edge point detector that screens edge points based on their spatial relationship to the edge. Additionally, we developed an edge attraction field extractor to compute the direction vector of the nearest edge for these points. This allows us to use the main directional tendencies of points around corner lines to verify true connections between corner points. By selecting diagonal connections, our method reduces false connections in the fully connected graph and addresses the sample imbalance, resulting in a complete roof wireframe. In experiments on synthetic datasets Point2Roof and the Real-world Building3D dataset, our method achieved superior performance. Notably, we improved Edge recall by 20 percentage points on Building3D. Our approach enhances all edge parameters on both datasets, outperforming existing wireframe reconstruction methods.