Long Lan
Long Lan

Long Lan

Associate Professor

College of Computer Science and Technology, National University of Defense Technology (NUDT)

State Key Laboratory of High-Performance Computing (HPCL)

lan19901@126.com
State Key Laboratory of High-Performance Computing, School of Computer, NUDT, Changsha, China
Google Scholar ↗GitHub ↗DBLP ↗arXiv ↗NUDT ↗

I am an Associate Professor with the College of Computer Science and Technology at the National University of Defense Technology (NUDT), Changsha, China.

I am affiliated with the State Key Laboratory of High-Performance Computing (HPCL). My research focuses on computer vision, machine learning, and efficient learning algorithms for real-world applications.

Research interests

Explore research →

My research focuses on designing efficient learning algorithms for real-world computer vision tasks.

01Visual Object Detection, Tracking and Re-identification02Few-shot Learning & Transfer Learning03Noisy Label Learning04Out-of-distribution Detection

Recent news

All updates →

Selected work

All publications →

On the Two Facets to Conquer Wild Out-of-distribution Detection

Research question
Auxiliary data collected in the wild often contain samples that share semantics with known classes. This contamination introduces incorrect ID/OOD labels and turns the wild OOD set into a mixture of ID and OOD distributions, which can mislead conventional outlier-exposure training.
Main findings
We address the problem from complementary instance and distribution perspectives. We dynamically estimate more reliable ID/OOD indicators and resample the wild data using the known ID distribution to reduce interference from the latent ID subset. The resulting unified framework is supported by theoretical analysis and extensive evaluation across multiple wild OOD settings.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2026
DOI ↗

Continuous Review and Timely Correction: Enhancing the Resistance to Noisy Labels via Self-Not-True and Class-Wise Distillation

Research question
Deep networks tend to learn clean samples first and gradually memorize incorrect labels. Early stopping alone cannot continuously prevent the accumulation of erroneous knowledge or recover useful information damaged by label noise.
Main findings
We introduce self-not-true distillation (SNTD), in which a model from an earlier training stage guides the current model while the true-class logit is masked, allowing non-true-class knowledge to support continuous review and timely correction. SNTD+ further selects teachers by class and dynamically adjusts the distillation weight for greater robustness under complex label noise.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2025
DOI ↗

From Concrete to Abstract: Multi-view Clustering on Relational Knowledge

Research question
Most multi-view clustering methods focus on concrete objects such as images and pay limited attention to abstract objects such as relational knowledge. Learning only basic consensus features can also overlook important sample-to-global correlations encoded in structural knowledge.
Main findings
We construct eight multi-view relational-knowledge datasets and propose RK-MVC. The framework captures coarse- and fine-grained sample-to-global correlations and fuses them with basic consensus features. Evaluation with nine representative backbones across seven aspects demonstrates broad compatibility and practical potential.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2025
DOI ↗

Join & collaborate

Prospective students & collaborators

We welcome motivated students and researchers who share our interests in computer vision and machine learning.

Research fit, application materials & availability

We welcome inquiries from students and researchers interested in computer vision, machine learning, and related areas. If your interests align with our work, please feel free to contact us by email.

Please briefly describe your academic or research background, the questions that interest you, and how you would like to participate. Include a CV or other relevant materials. Prospective students should indicate the program and intended start date; potential collaborators may outline the topic they would like to explore together.

We respond based on current research plans, supervision capacity, and alignment of interests. You are also welcome to stay in touch and follow future updates.