Qinghe Liu

Team Leader (5 Members)

06/2022 - 06/2024

Research on Segmentation and Recognition of Three-Dimensional Fruit Traits Based on Point Cloud Deep Learning

• Used the MVS algorithm and depth camera to produce strawberry dataset in complex scenes;
• Adopted density-based method to improve the standard point cloud deep learning framework;
• Designed FSDnet to improve the segmentation accuracy of the standard point cloud deep learning framework in the agricultural picking activities.
• Led and participated in the production of datasets, the review and summary of related papers and technical codes, the planning of the project process, the writing and debugging of project codes, the design and implementation of project experiments, the writing of project paper, and the application and completion of the project.
• Noticed the development trends and technological changes in the field of visualization technologies;
• Got clear design idea for training and experiments of deep learning and developed ability to solve complicated problems in the process;
• Deeply understood the methods of multi-GPU distributed training; clearly knew the process for cropping, denoising, integration, dicing, and encapsulation of the dataset.

Member (5 members)

06/2023 - 06/2024

Design and Research of Point Cloud Neural Network Based on High-Order Feature Interaction

• With the help of Image HorNet, utilized the self-attention mechanism to design the Point HorNet in the field of point cloud segmentation.
• Achieved mloU 73% (ranked 8th) in the Stanford 3D Indoor Scene Dataset (S3DIS) 6-fold cross-validation.
• As a member, mainly took the responsibilities of code design and validity verification for the migration from the initial Image HorNet to the Point HorNet.

Writer (Graduation Thesis)

12/2023 - 06/2024

Comprehensive Analysis and Research on Point Cloud Grouping Sampling Algorithms

• Analyzed that the bottleneck currently restricting the application of point cloud neural network is the excessively high time complexity of the point cloud grouping and sampling algorithm;
• Designed grid sampling and grouping method and improved the efficiency by 35 times.
• Collected and analyzed the information about the research status of point cloud deep learning;
• Used CUDA to implement grid sampling and grouping algorithm and conducted mosaic comparison experiments;
• Designed a lightweight network GPFEnet.

Member (5 members)

06/2021 - 09/2022

Automatic Fruit Recognition Based on 3D Point Cloud

• Applied MVS algorithm to create point cloud datasets of fruits such as apple and strawberry;
• Used standard point cloud deep learning frameworks (PointNet ++, PAConv) for segmentation and recognition.
• As the key designer and implementer of the project experiments, reproduced and compared multiple project codes such as PointNet, PointNet ++, PACONV, and Point Transformer on agricultural datasets.
• Became familiar with the research and improvement directions of point cloud deep learning.

Paper Title: FSDNET: A Features Spreading Net with Density for 3D Segmentation in Agriculture
Magazine: Computers and Electronics in Agriculture; Impact Factor: 8.3
Date of Publication: May, 2024
ISSN 0168-1699

Title: GPFEnet: A Lightweight Grid Parallel Feature Extraction Net for 3D Segmentation in Agriculture
Accepted by: International Joint Conference on Neural Networks (IJCNN)
Date of Conference: March, 2025
Conference Location: Rome, Italy

Title: Point HorNet: Higher Order Spatial Interaction Network for Point Clouds
Accepted by: International Conference on Virtual Reality (ICVR)
Date of Conference: July, 2024
Conference Location: Bournemouth, UK