Peiwen Sun

I am a Researcher at BUPT, focused on machine learning for emotion/sentiment understanding and multimodal understanding. After completing my undergraduate studies at BUPT, I was recommended to pursue a master's degree in the Department of Artificial Intelligence at the same institution. I have worked closely with Di Hu from Renmin University of China and Wei Xue from Hong Kong University of Science and Technology. I am heading to my PhD in the future.

I used to be an intern at Megvii and Tencent to do research on multimodal understanding and generation.

GitHub  /  Page

🎉7/16/2024: 2 papers are accepted by ACM MM 2024. See you in Melbourne.

🎉7/1/2024: 3 papers are accepted by ECCV 2024. See you in Milan.

2025: ICLR, ICASSP, COLING

2024: ICLR, CVPR, ECCV

My research focuses on audio-visual learning and multi-modal machine learning, with the downstream tasks of person recognition and emotion/sentiment recognition techniques for video recognition, including the use of sound and images to learn better representations.

The pioneering work on language-driven spatial audio generation for controllable and immersive stereo audio.

Unveiling and mitigating bias caused by real-world inherent preferences and distributions in Audio Visual Segmentation.

A novel task of referring segmenation by audio, visual and temporal information.

The ability of LLM can be used to solve segmenation preference on Audio Visual Segmention.

A novel audio-visual strategy in person verification that considers connections between time series from a generative perspective.

This paper reconsiders the information fused in the multimodal case from a bionics perspective and proposes a simple, plug-and-play, attention module for vanilla fusion based on fundamental information theory and uncertainty theory.

I have been exposed to a variety of research directions such as few-shot learning, medical image processing, mathematical modelling etc.

A large-scale zero-shot speech synthesis system with approximately 5% of the inference time compared with previous work.

A two-stage training strategy for AVS, which decomposes the AVSS task into two simple subtasks from localization to semantic understanding, which to achieve step-by-step global optimization.

In this paper, we propose a series of strategies to train a better backbone for few-shot learning.

In this paper, the popular semantic segmentation network is firstly applied in thyroid nodule classification in ultrasound images.

A modeling scheme for the layout and number of drones to extinguish and monitor wildfires in southern Australia.