Recently, the Big Data and Artificial Intelligence team from the School of Artificial Intelligence published a research paper titled CVKD-UDA: Cross-View Knowledge Distillation for 3D Unsupervised Domain Adaptive Segmentation in IEEE Transactions on Multimedia (TMM). TMM is a top-tier international journal in the fields of computer vision and multimedia under the Institute of Electrical and Electronics Engineers (IEEE). It is a Class A international journal recommended by the China Computer Federation (CCF) and a CAS (Chinese Academy of Sciences) Q1 journal. The first author of the paper is Dr. Yuan Zhimin from Nanyang Normal University, the corresponding author is Professor Cheng Ming from Xiamen University, and NYNU is the first corresponding affiliation.

This method identifies voxel size as a key design factor for constructing cross-domain similar representations. By leveraging the complementary information between different voxel views, it balances the transferability and discriminability of the pre-trained model. Experimental results demonstrate that the CVKD-UDA method achieves excellent performance on two LiDAR point cloud unsupervised domain adaptive segmentation datasets, providing a concise and effective new approach for 3D point cloud cross-domain semantic segmentation.

The team also published a research article titled CIMD: Cognitive Inspired Multi-perspective Descriptions for EEG-Image Alignment in the IEEE International Conference on Multimedia and Expo (ICME). Sponsored by the IEEE, ICME is one of the most representative international academic conferences in the fields of computer graphics and multimedia and is a Class B international academic conference recommended by the CCF. The first author of this paper is Dr. Li Jingjing from NYNU, the second author is Liu Xinqi, a master’s student at NYNU, the corresponding author is Dr. Huang Xin from NYNU, and NYNU is the first corresponding affiliation.

Electroencephalography (EEG)-image retrieval is an important task in the brain-computer interface field for achieving semantic association between neural signals and visual stimuli. This paper proposes a fine-grained alignment framework using a cognitively inspired multi-perspective description. It leverages multimodal large models to generate cognitively inspired text descriptions from global, color, and emotional perspectives. Then, using a cognition-driven multimodal alignment module, these multi-perspective text descriptions guide the fine-tuning of a pre-trained model to construct a unified shared embedding space for EEG, image, and text. This achieves more refined cross-modal semantic alignment, constrains feature drift during the fine-tuning process, and enables effective cross-modal EEG-image alignment.  

Written & Photo by: Yuan Zhimin & Yang Shiwei

Source: NYNU Academic Activities (Chinese)

https://www.nynu.edu.cn/info/1048/31568.htm