Ph.D. Researcher · NUDT 博士研究生 · 国防科技大学
College of Systems Engineering · Laboratory for Big Data and Decision 系统工程学院 · 大数据与决策实验室
I am a Ph.D. researcher at the National University of Defense Technology (NUDT), working at the College of Systems Engineering and the Laboratory for Big Data and Decision. My research asks how a visual model can keep the information that matters when the world shows it two very different views of the same thing — a visible image and a thermal one, or two crops of a scene under rotation.
我是国防科技大学博士研究生，在系统工程学院与大数据与决策实验室开展研究。我的研究关心这样一个问题：当世界把同一事物的两种不同视角交给模型——同一场景的可见光图像与红外图像，或者同一场景在旋转下的两个视图——视觉模型怎样才能保留真正有用的信息。
My work spans multi-modal representation learning and infrared–visible fusion, minimal contrastive learning, contrastive manifold autoencoders for latent diffusion, and equivariant 360° vision. I am a co-first author of CMAE (IEEE ICME 2026, Oral) and a co-author of EquiSSL (ACM TOG, SIGGRAPH Asia 2026), DAGA (NeurIPS 2026), CORE (IEEE TNNLS), SDI (Pattern Recognition), HTNet (Neural Networks) and TriCM-Net (PRCV 2026). I am also a co-inventor of two granted national invention patents.
我的研究涵盖多模态表征学习与红外—可见光融合、极简对比学习、面向潜在扩散的对比流形自编码，以及等变 360° 视觉。我是 CMAE（IEEE ICME 2026，Oral）的共同第一作者，也是 EquiSSL（ACM TOG，SIGGRAPH Asia 2026）、DAGA（NeurIPS 2026）、CORE（IEEE TNNLS）、SDI（Pattern Recognition）、HTNet（Neural Networks）与 TriCM-Net（PRCV 2026）的合作者，并作为共同发明人参与了两项已授权的国家发明专利。
Latest最新动态
Our paper EquiSSL is accepted by ACM Transactions on Graphics (SIGGRAPH Asia 2026 Journal Track).
论文 EquiSSL 被 ACM Transactions on Graphics 接收（SIGGRAPH Asia 2026 期刊轨）。
CORE — Less Is More — is published in IEEE Transactions on Neural Networks and Learning Systems.
CORE（Less Is More）正式发表于 IEEE Transactions on Neural Networks and Learning Systems。
DAGA is accepted to NeurIPS 2026 — dynamic attention-guided adaptation for frozen self-supervised vision transformers.
DAGA 被 NeurIPS 2026 接收——面向冻结自监督视觉 Transformer 的动态注意力引导适配。
The co-first-author paper CMAE is accepted for an Oral presentation at IEEE ICME 2026 — a contrastive manifold autoencoder for efficient latent diffusion.
共同第一作者论文 CMAE 被 IEEE ICME 2026 接收并作 Oral 报告——面向高效潜在扩散的对比流形自编码。
Two national invention patents are granted by CNIPA — CN119540575B and CN119580024B.
两项国家发明专利获国家知识产权局授权——CN119540575B 与 CN119580024B。
ACM Transactions on Graphics, Vol. 45, No. 6, Article 242
TL;DR — Removes gauge dependence from spherical attention with a parameter-free Reynolds average, cutting the 90° rotation drop from 53% to 0.8%.
一句话总结——以无参数的 Reynolds 平均消除球面注意力中的规范依赖，将 90° 旋转下的性能下降从 53% 降至 0.8%。
IEEE International Conference on Multimedia and Expo (ICME 2026)
TL;DR — Replaces VAE latents with contrastive manifold autoencoding, preserving high-frequency detail and accelerating diffusion sampling by 2.1×.
一句话总结——以对比流形自编码替代 VAE 潜空间，保留高频细节，并将扩散采样加速约 2.1 倍。
Advances in Neural Information Processing Systems (NeurIPS 2026)
TL;DR — A dynamic, parameter-efficient adapter for frozen self-supervised vision transformers, improving downstream transfer.
一句话总结——面向冻结自监督视觉 Transformer 的动态参数高效适配器，提升下游迁移性能。
IEEE Transactions on Neural Networks and Learning Systems
TL;DR — Deconstructs contrastive learning to its minimal sufficient components; 42% less compute with higher linear-probe accuracy.
一句话总结——把对比学习拆解到最小充分组件；计算量减少 42%，线性探测精度反而更高。
* corresponding author通讯作者 · † equal contribution同等贡献
Ph.D. researcher, NUDT · College of Systems Engineering; Laboratory for Big Data and Decision.
博士研究生，国防科技大学 · 系统工程学院；大数据与决策实验室。
Co-first author · CMAE, IEEE ICME 2026 (Oral).
共同第一作者 · CMAE，IEEE ICME 2026（Oral）。
Co-author · EquiSSL (ACM TOG), DAGA (NeurIPS), CORE (IEEE TNNLS), SDI (Pattern Recognition), HTNet (Neural Networks), TriCM-Net (PRCV).
合作者 · EquiSSL（ACM TOG）、DAGA（NeurIPS）、CORE（IEEE TNNLS）、SDI（Pattern Recognition）、HTNet（Neural Networks）、TriCM-Net（PRCV）。
Co-inventor · two granted CNIPA national invention patents (CN119540575B, CN119580024B).
共同发明人 · 两项国家发明专利获授权（CN119540575B、CN119580024B）。
Collaboration合作
For research collaborations, discussions of a paper, or questions about the released code, write to me at [email protected]. 无论是研究合作、论文讨论，还是关于已公开代码的问题，都欢迎写信至 [email protected]。