Camera-Incremental Person Re-Identification via Geometric Representation Evolution

Mang Ye, Shipeng Zhang, Wuxuan Shi, Zhijie Lu, Xingyu Gao, He Li

• A camera-incremental person re-identification framework that preserves global feature geometry while consolidating recurring identities across newly deployed cameras.

PRISM: Spectral Pruning and Reconstruction for Parameter-Efficient Model Merging of MLLMs

Wuxuan Shi, Haotian Chen, He Li, Qiang Yang, Mang Ye

• A data-free framework for parameter-efficient MLLM model merging that spectrally prunes low-energy noise and orthogonally reconstructs high-energy task updates.

Towards Realistic Lifelong Re-identification: Identity Recurrence with Changing Clothes

Wuxuan Shi*, Zhijie Lu*, He Li, Mang Ye

Code

• Introducing a realistic lifelong cloth-changing person re-identification setting with recurring identities, identity-clothing disentanglement, and dynamic identity-anchor alignment.

Prototype-guided knowledge propagation with adaptive learning for lifelong person re-identification

Zhijie Lu*, Wuxuan Shi*, He Li, Mang Ye

Code

• A lifelong person re-identification method with prototype improved triplet loss and dynamic weight updates.

Synthetic Data is an Elegant GIFT for Continual Vision-Language Models

Bin Wu*, Wuxuan Shi*, Jinqiao Wang, Mang Ye

Code

• Using synthetic data to help VLMs achieve a better balance between maintaining generalization capabilities and learning downstream tasks.

Prospective Representation Learning for Non-Exemplar Class-Incremental Learning

Wuxuan Shi, Mang Ye

• Prepare the model for handling conflicts of old and new classes in advance.

Prototype Reminiscence and Augmented Asymmetric Knowledge Aggregation for Non-Exemplar Class-Incremental Learning

Wuxuan Shi, Mang Ye

Code

• Improving class incremental learning without storing old task exemplars.