Xin Wen / 温鑫


I am a final year Ph.D. Candidate at HKU CVMI Lab, advised by Prof. Xiaojuan Qi. Before that, I got my B.Eng. degree in Computer Science (minor in Mathematics) at Tongji University.

I work on (self-supervised) representation learning, with interest in learning robust, compositional, and generalizable visual representations from uncurated data with minimal human intervention, and its applications to perception & generation tasks.

I will join Meta FAIR as a Research Scientist Intern in June 2025, and will be on job market for Research Scientist and Postdoc positions starting December 2025.

Email  /  Github  /  Google Scholar  /  Twitter


profile photo
Selected / All   Publications
"Principal Components" Enable A New Language of Images
Xin Wen*, Bingchen Zhao*, Ismail Elezi, Jiankang Deng, Xiaojuan Qi
Preprint, 2025
Project Page / ArXiv / Code

We introduce a novel visual tokenization framework that embeds a provable PCA-like structure into the latent token space, enabling 1D coarse-to-fine tokenization.

A Data-Centric Revisit of Pre-Trained Vision Models for Robot Learning
Xin Wen, Bingchen Zhao, Yilun Chen, Jiangmiao Pang, Xiaojuan Qi
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025
ArXiv / Code

We designed SlotMIM, a method that induces object-centric representations from non-object-centric images, which we find facilitates robot learning.

Learning from Neighbors: Category Extrapolation for Long-Tail Learning
Shizhen Zhao, Xin Wen, Jiahui Liu, Chuofan Ma, Chunfeng Yuan, Xiaojuan Qi
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025
ArXiv

We find extrapolating tail classes with novel classes that share similar semantics with tail classes significantly improves long-tail recognition.

What Makes CLIP More Robust to Long-Tailed Pre-Training Data? A Controlled Study for Transferable Insights
Xin Wen, Bingchen Zhao, Yilun Chen, Jiangmiao Pang, Xiaojuan Qi
Conference on Neural Information Processing Systems (NeurIPS), 2024
ArXiv / Code / Slides / Poster

We find CLIP to be relatively robust to pre-training data imbalance, design and conduct controlled experiments to identify the underlying mechanisms and provide insights for recognition and SSL models.

Can OOD Object Detectors Learn from Foundation Models?
Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, Xiaojuan Qi
European Conference on Computer Vision (ECCV), 2024
ArXiv / Code

We introduce an automatic data curation process that leverages foundation models as tools to harvest meaningful data from text-to-image generation models for OOD object detection.

Classes Are Not Equal: An Empirical Study on Image Recognition Fairness
Jiequan Cui, Beier Zhu, Xin Wen, Xiaojuan Qi, Bei Yu, Hanwang Zhang
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024
ArXiv / Code

We show that models can still have extremely biased behaviors when trained on balanced ImageNet, investigate the resons behind, and provide some workarounds.

What If the TV Was Off? Examining Counterfactual Reasoning Abilities of Multi-modal Language Models
Letian Zhang, Xiaotong Zhai, Zhongkai Zhao, Yongshuo Zong, Xin Wen, Bingchen Zhao
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024
Project Page / ArXiv / Code

We build a counterfactual visual question answering benchmark, and show that strong Vision-Language Models, even GPT-4, cannot handle them very well.

Towards Large-scale 3D Representation Learning with Multi-dataset Point Prompt Training
Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, Hengshuang Zhao
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024
ArXiv / Code

We investigate the techniques to enable 3D representation learning at unprecedented data scale.

CoDet: Co-Occurrence Guided Region-Word Alignment for Open-Vocabulary Object Detection
Chuofan Ma, Yi Jiang*, Xin Wen*, Zehuan Yuan, Xiaojuan Qi
Conference on Neural Information Processing Systems (NeurIPS), 2023
Project Page / ArXiv / Code

We bridge the gap between vision & language spaces by reformulating region-word alignment as co-occurring object discovery, and images mention a shared concept in their captions are grouped together.

Parametric Classification for Generalized Category Discovery: A Baseline Study
Xin Wen*, Bingchen Zhao*, Xiaojuan Qi
IEEE International Conference on Computer Vision (ICCV), 2023
Project Page / ArXiv / Code / Slides / Poster

We revisit the reason that makes previous parametric classifiers fail to recognise new classes for GCD, identify the prediction biases between and within seen and novel classes as the key issue, and propose a simple yet strong framework that addresses these limitations and achieves state-of-the-art performance in this field.

Learning Semi-supervised Gaussian Mixture Models for Generalized Category Discovery
Bingchen Zhao, Xin Wen, Kai Han
IEEE International Conference on Computer Vision (ICCV), 2023
ArXiv / Code

We tackle GCD without knowing the class number as a-priori, propose a semi-supervised variant of GMM with stochastic splitting and merging to dynamically determine prototypes, and leverage prototpyical contrastive learning for representation learning on partially labelled data.

Masked Scene Contrast: A Scalable Framework for Unsupervised 3D Representation Learning
Xiaoyang Wu, Xin Wen, Xihui Liu, Hengshuang Zhao
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023
ArXiv / Code

We propose the Masked Scene Contrast (MSC) framework for unsupervised 3D representation learning, which efficiently generates contrastive views directly on scene-level point clouds and enables large-scale 3D pre-training across multiple datasets.

Self-Supervised Visual Representation Learning with Semantic Grouping
Xin Wen, Bingchen Zhao, Anlin Zheng, Xiangyu Zhang, Xiaojuan Qi
Conference on Neural Information Processing Systems (NeurIPS), 2022
Project Page / ArXiv / Code / Slides / Poster

We show that object discovery can be learned jointly with the representations from scratch on real-world scene-centric data, which leads to strong transfer learning results in various downstream tasks.

Temporal Context Aggregation for Video Retrieval with Contrastive Learning
Jie Shao*, Xin Wen*, Bingchen Zhao, Xiangyang Xue
IEEE Winter Conference on Applications of Computer Vision (WACV), 2021
ArXiv / Code / Slides

We present a contrastive learning-based video representation learning framework that adopts long-range temporal information between frame-level features using self-attention.

Distilling Visual Priors from Self-Supervised Learning
Bingchen Zhao, Xin Wen
European Conference on Computer Vision (ECCV) VIPriors Workshop, 2020
ArXiv / Code / Slides

We leverage self-supervised learning and knowledge distillation to improve the generalizability of CNN models for image classification under the data-deficient setting.

Invited Talks

  • CCVL Lab, Johns Hopkins University: "Self-Supervised Visual Representation Learning with Semantic Grouping", Nov. 2022.
  • Awards

    NeurIPS 2022 Scholar Award Oct. 2022
    Outstanding Graduates of Shanghai June 2021
    2nd place in the ECCV 2020 Workshop VIPriors Image Classification Challenge July 2020
    Qidi Scholarship of Tongji University (top 1%) June 2020
    Regional Champion (China) of the Covestro International Data Science Hackathon Nov. 2019
    Silver Medal of the 43rd ACM International Collegiate Programming Contest (ICPC) Asia-East Continent Final Dec. 2018
    Academic Services

    Reviewer for TPAMI, IJCV, NeurIPS, ICLR, ICML, CVPR, ICCV, ECCV, WACV, CVinW, and OOD-CV.


    Template gratefully stolen from here.