Xin Jin 金鑫

I am an Assistant Professor at the College of Info Science & Technology, Eastern Institute of Technology (EIT), Ningbo where I cooperated with professor Wenjun Zeng (IEEE Fellow). Our group is to engage in cutting-edge research in computer vision and multimedia, recently focusing on spatial and embodied AI.

Previously, I was a Visiting Scholar of Learning and Vision (LV) Lab at the National University of Singapore where I was guided by professor Xinchao Wang, professor Jiashi Feng and professor Shuicheng Yan. I received Ph.D. degree from University of Science and Technology of China (USTC), under the supervision of Zhibo Chen. From Jan. 2019 to Jul. 2020, I also worked at Intelligent Multimedia Group (IMG) in MSRA under the supervision of Cuiling Lan. From Sep. 2018 to Jan. 2019, I worked at KDDI Research, Inc. in Japan under the supervision of Jianfeng Xu.

If you are highly creative, have top research/coding skill and interested in joining us, please do not hesitate to send me (jinxin@eitech.edu.cn) your CV.

Email: jinxin@eitech.edu.cn  /  Google Scholar  /  Github

Recently, in ICCV 2025, we organized the 1st International Workshop and Challenge on Disentangled Representation Learning for Controllable Generation. In CVPR 2024 and ECCV 2024, we organized two tutorial sessions related to “Visual Disentanglement and Compositionality”.

We recast the vision–language–action model as a perception–prediction–action model and make the model explicitly predict a compact set of dynamic, spatial and high- level semantic information, supplying concise yet comprehensive look-ahead cues for planning.

This survey outlines the concept of Behavior Foundation Models (BFMs) for humanoid whole-body control, detailing large-scale pre-training workflows that enable reusable motion primitives and zero-shot adaptation across diverse tasks. It highlights challenges and potential in applying BFMs to real-world systems and curates a growing repository of related works.

We introduce the concept of semantic orientation, representing the object orientation condition on open vocabulary language.

We tackle the problem of how to build and benchmark a large motion model with video action datasets and disentangled rule-based annotations.

ULTHO formulates hyperparameter optimization for deep RL as a multi-armed bandit over clustered hyperparameter configurations, enabling efficient early pruning of underperforming runs and focusing compute on promising settings. Evaluated on benchmarks like ALE, Procgen, MiniGrid, and PyBullet, ULTHO achieves near‐optimal performance using just a fraction of the usual computational budget.

Disentangled World Models (DisWM) introduce a model-based RL framework that leverages offline pretraining on distracting videos with disentanglement regularization and offline-to-online latent distillation. This enables transfer of semantic knowledge for improved sample efficiency and generalization in downstream reinforcement learning tasks in visually varied environments.

we propose Hybrid-depth, a novel framework that systematically integrates foundation models (CLIP and DINO) to extract visual priors and acquire sufficient contextual information for self-supervised depth estimation methods.

BridgeAD introduces multi-step motion and planning queries that integrate historical prediction into both perception and planning modules, thereby “bridging” the past and future. With this structure, the model unifies historical insight with current perception and future trajectory planning—achieving state-of-the-art performance on nuScenes in open- and closed-loop settings.

This paper systematically investigates multi-layer visual feature fusion in Multimodal Large Language Models, analyzing optimal layer selection and four fusion strategies. Results show that external direct fusion of features from distinct encoder stages consistently offers the best generalization and stability. Their code is publicly released.

UniScene proposes the first occupancy‑centric hierarchical framework that consecutively generates semantic occupancy, multi-view video, and LiDAR data from coarse BEV layouts, using Gaussian‑based joint rendering and prior‑guided sparse modeling. This approach significantly outperforms previous methods in fidelity and versatility across all three modalities, benefiting downstream perception tasks :contentReference[oaicite:0]{index=0}.

We present LS-Imagine, a model-based RL framework that constructs a long short-term world model combining step-by-step and jumpy transitions, guided by affordance maps to enable goal-conditioned exploration in open-world environments. This approach significantly improves exploration efficiency and sample efficiency in complex, high-dimensional tasks.

This work presents DisCo, a framework that leverages scene graphs as structured conditions to generate complex images by disentangling layouts and semantics via a Semantics-Layout VAE and composing them with a diffusion-based Compositional Masked Attention, while enabling isolated graph-guided editing through a Multi-Layered Sampler—achieving state-of-the-art generalization across diverse scene complexities.

This work introduces a bidirectional weighted graph framework that integrates β‑VAE to extract latent factors and leverages multimodal LLMs to detect and weight semantic correlations, leading to fine‐grained, interpretable disentanglement and strong reconstruction performance.

This paper presents CoWorld—a model-based reinforcement learning framework that bridges offline and online domains by leveraging auxiliary simulators as “test beds.” It aligns latent state and reward distributions across domains and introduces min‑max value constraints to mitigate overestimation, yielding substantial improvements over existing offline visual RL methods.

In this work, by rethinking the inherent relationship between the face of target identity and its variants, we introduce a new pipeline of Generalized Manifold Adversarial Attack (GMAA) to achieve a better attack performance by expanding the attack range.

In this work, we propose a new efficient tuning approach for VLMs named Task Residual Tuning (TaskRes), which performs directly on the text-based classifier and explicitly decouples the prior knowledge of the pre-trained models and new knowledge regarding a target task.

In this work, we are the first to propose a novel training strategy for image restoration from the causality perspective, to improve the generalization ability of DNNs for unknown degradations.

In this work, we resort to data mixing to establish a deliberated domain bridging (DDB) for domain adaptive semantic segmentation. The joint distributions of source and target domains are aligned and interacted with each other in the intermediate space.

In this paper, we attempt to learn a kind of omnipotent feature that is both general (for AI tasks) and compact (for compression) for Image Coding for Machines (ICM). Considering self-supervised learning (SSL) improves feature generalization, we integrate it with the compression task to learn such features.

In this paper, we explore a novel learning scheme, termed as Learning wIth Recoverable Forgetting (LIRF), that explicitly handles the task- or sample-specific knowledge removal and recovery.

In this paper, we make attempt to the large-scale Unsupervised ReID and propose a “small data for big task” paradigm dubbed Meta Clustering Learning (MCL), which our method significantly saves computational cost while achieving a comparable or even better performance compared to prior works.

We design an Unsupervised Pre-training framework for ReID based on the contrastive learning (CL) pipeline, dubbed UP-ReID.

We focus on handling well the Cloth-Changing ReID problem under a more challenging setting, i.e., just from a single image, which enables high-efficiency and latency-free pedestrian identify for real-time surveillance applications.

We address the adversarial-based DA problem from a different perspective and design a simple yet effective adversarial paradigm in the form of a discriminator-free adversarial learning network (DALN), wherein the category classifier is reused as a discriminator.

We propose the Dual Prior Learning (DPL) method for blind image restoration by taking both image and distortion priors into account. DPL goes beyond DIP (deep image prior) by considering an additional step to explicitly learn the blended distortion prior.

We design a novel Style Normalization and Restitution module (SNR) to simultaneously ensure both high generalization and discrimination capability of the networks, and evaluate it on multiple vision tasks of classification, detection, segmentation, etc.

We propose an efficient optimization strategy named Re-enforceable Adversarial Domain Adaptation (RADA) which aims to re-energize the domain discriminator during the training by using dynamic domain labels.

This paper proposes a hybrid framework, Dense Interaction Learning (DenseIL), that takes the principal advantages of both CNN-based and Attention-based architectures to tackle video-based person re-ID difficulties.

The key to solving this more challenging real image super-resolution (RealSR) problem lies in learning feature representations that are both informative and content-aware. We propose an Omni-frequency Region-adaptive Network (OR-Net), here we call features of all low, middle and high frequencies omni-frequency features.

We introduce a global distance-distributions separation (GDS) constraint over the two distributions to encourage the clear separation of positive and negative samples from a global view.

We introduce the concept of Disentangled Feature Learning to achieve the feature-level divide-and-conquer of hybrid distortions for low-level enhancement.

We propose a simple yet effective Style Normalization and Restitution (SNR) module. Specifically, we filter out style variations (eg, illumination, color contrast) by Instance Normalization (IN). However, such a process inevitably removes discriminative information. We propose to distill identity-relevant feature from the removed information and restitute it to the network to ensure high discrimination.

We propose an effective Relation-Aware Global Attention (RGA) module which captures the global structural information for better attention learning.

We build a Semantics Aligning Network (SAN) which consists of a base network as encoder (SA-Enc) for re-ID, and a decoder (SA-Dec) for reconstructing the densely semantics aligned full texture image.

We propose exploiting the multi-shots of the same identity to guide the feature learning of each individual image. Specifically, we design an Uncertainty-aware Multi-shot Teacher-Student (UMTS) Network.

We show that the mean and variance shifts caused by full-spatial FN limit the image inpainting network training and we propose a spatial region-wise normalization named Region Normalization (RN) to overcome the limitation.

Feel free to steal this website's source code.