Weixin Li

Jiankai Li, Yunhong Wang, Weixin Li

Scene Graph Generation (SGG) plays a pivotal role in downstream vision-language tasks. Existing SGG methods typically suffer from poor compositional generalizations on unseen triplets. They are generally trained on incompletely annotated scene graphs that contain dominant triplets and tend to bias toward these seen triplets during inference. To address this issue, we propose a Triplet Calibration and Reduction (T-CAR) framework in this paper. In our framework, a triplet calibration loss is first presented to regularize the representations of diverse triplets and to simultaneously excavate the unseen triplets in incompletely annotated training scene graphs. Moreover, the unseen space of scene graphs is usually several times larger than the seen space since it contains a huge number of unrealistic compositions. Thus, we propose an unseen space reduction loss to shift the attention of excavation to reasonable unseen compositions to facilitate the model training. Finally, we propose a contextual encoder to improve the compositional generalizations of unseen triplets by explicitly modeling the relative spatial relations between subjects and objects. Extensive experiments show that our approach achieves consistent improvements for zero-shot SGG over state-of-the-art methods. The code is available at https://github.com/jkli1998/T-CAR.

Yutong Zhang, Jiaxin Chen, Honglin Chen et al.

Memory-efficient transfer learning (METL) approaches have recently achieved promising performance in adapting pre-trained models to downstream tasks. They avoid applying gradient backpropagation in large backbones, thus significantly reducing the number of trainable parameters and high memory consumption during fine-tuning. However, since they typically employ a lightweight and learnable side network, these methods inevitably introduce additional memory and time overhead during inference, which contradicts the ultimate goal of efficient transfer learning. To address the above issue, we propose a novel approach dubbed Masked Dual Path Distillation (MDPD) to accelerate inference while retaining parameter and memory efficiency in fine-tuning with fading side networks. Specifically, MDPD develops a framework that enhances the performance by mutually distilling the frozen backbones and learnable side networks in fine-tuning, and discard the side network during inference without sacrificing accuracy. Moreover, we design a novel feature-based knowledge distillation method for the encoder structure with multiple layers. Extensive experiments on distinct backbones across vision/language-only and vision-and-language tasks demonstrate that our method not only accelerates inference by at least 25.2\% while keeping parameter and memory consumption comparable, but also remarkably promotes the accuracy compared to SOTA approaches. The source code is available at https://github.com/Zhang-VKk/MDPD.

Jinyuan Liu, Yang Wang, Zeyu Zhao et al.

Reasoning video object segmentation predicts pixel-level masks in videos from natural-language queries that may involve implicit and temporally grounded references. However, existing methods are developed and evaluated in an offline regime, where the entire video is available at inference time and future frames can be exploited for retrospective disambiguation, deviating from real-world deployments that require strictly causal, frame-by-frame decisions. We study Online Reasoning Video Object Segmentation (ORVOS), where models must incrementally interpret queries using only past and current frames without revisiting previous predictions, while handling referent shifts as events unfold. To support evaluation, we introduce ORVOSB, a benchmark with frame-level causal annotations and referent-shift labels, comprising 210 videos, 12,907 annotated frames, and 512 queries across five reasoning categories. We further propose a baseline with continually-updated segmentation prompts and a structured temporal token reservoir for long-horizon reasoning under bounded computation. Experiments show that existing methods struggle under strict causality and referent shifts, while our baseline establishes a strong foundation for future research.

Xiang Lin, Weixin Li, Shu Guo et al.

Pre-trained Vision-Language Models (VLMs), \textit{e.g.} CLIP, have become essential tools in multimodal transfer learning. However, fine-tuning VLMs in few-shot scenarios poses significant challenges in balancing task-specific adaptation and generalization in the obtained model. Meanwhile, current researches have predominantly focused on prompt-based adaptation methods, leaving adapter-based approaches underexplored and revealing notable performance gaps. To address these challenges, we introduce a novel Reconstruction-based Multimodal Adapter (RMAdapter), which leverages a dual-branch architecture. Unlike conventional single-branch adapters, RMAdapter consists of: (1) an adaptation branch that injects task-specific knowledge through parameter-efficient fine-tuning, and (2) a reconstruction branch that preserves general knowledge by reconstructing latent space features back into the original feature space. This design facilitates a dynamic balance between general and task-specific knowledge. Importantly, although RMAdapter introduces an additional reconstruction branch, it is carefully optimized to remain lightweight. By computing reconstruction loss locally at each layer and sharing projection modules, the overall computational overhead is kept minimal. A consistency constraint is also incorporated to better regulate the trade-off between discriminability and generalization. We comprehensively evaluate the effectiveness of RMAdapter on three representative tasks: generalization to new categories, generalization to new target datasets, and domain generalization. Without relying on data augmentation or duplicate prompt designs, our RMAdapter consistently outperforms state-of-the-art approaches across all evaluation metrics.

Zixing Lei, Changxing Liu, Yichen Xiong et al.

Vision-language-action (VLA) models are effective robot action executors, but they remain limited on long-horizon tasks due to the dual burden of extended closed-loop planning and diverse physical operations. We therefore propose VLAs-as-Tools, a strategy that distributes this burden across a high-level vision language model (VLM) agent for temporal reasoning and a family of specialized VLA tools for diverse local physical operations. The VLM handles scene analysis, global planning, and recovery, while each VLA tool executes a bounded subtask. To tightly couple agent planning with VLA tool execution in long-horizon tasks, we introduce a VLA tool-family interface that exposes explicit tool selection and in-execution progress feedback, enabling efficient event-triggered agent replanning without continuous agent polling. To obtain diverse specialized VLA tools that faithfully follow agent invocations, we further propose Tool-Aligned Post-Training (TAPT), which constructs invocation-aligned training units for instruction following and adopts tool-family residual adapters for efficient tool specialization. Experiments show that VLAs-as-Tools improves the success rate of $π_{0.5}$ by 4.8 points on LIBERO-Long and 23.1 points on RoboTwin, and further enhances invocation fidelity by 15.0 points as measured by Non-biased Rate. Code will be released.

Xin Lu, Rui Li, Xun Huang et al.

Embodied Question Answering (EQA) has traditionally been evaluated in temporally stable environments where visual evidence can be accumulated reliably. However, in dynamic, human-populated scenes, human activities and occlusions introduce significant perceptual non-stationarity: task-relevant cues are transient and view-dependent, while a store-then-retrieve strategy over-accumulates redundant evidence and increases inference cost. This setting exposes two practical challenges for EQA agents: resolving ambiguity caused by viewpoint-dependent occlusions, and maintaining compact yet up-to-date evidence for efficient inference. To enable systematic study of this setting, we introduce DynHiL-EQA, a human-in-the-loop EQA dataset with two subsets: a Dynamic subset featuring human activities and temporal changes, and a Static subset with temporally stable observations. To address the above challenges, we present DIVRR (Dynamic-Informed View Refinement and Relevance-guided Adaptive Memory Selection), a training-free framework that couples relevance-guided view refinement with selective memory admission. By verifying ambiguous observations before committing them and retaining only informative evidence, DIVRR improves robustness under occlusions while preserving fast inference with compact memory. Extensive experiments on DynHiL-EQA and the established HM-EQA dataset demonstrate that DIVRR consistently improves over existing baselines in both dynamic and static settings while maintaining high inference efficiency.

Jiankai Li, Yunhong Wang, Xiefan Guo et al.

Scene Graph Generation (SGG) aims to identify entities and predict the relationship triplets \textit{\textless subject, predicate, object\textgreater } in visual scenes. Given the prevalence of large visual variations of subject-object pairs even in the same predicate, it can be quite challenging to model and refine predicate representations directly across such pairs, which is however a common strategy adopted by most existing SGG methods. We observe that visual variations within the identical triplet are relatively small and certain relation cues are shared in the same type of triplet, which can potentially facilitate the relation learning in SGG. Moreover, for the long-tail problem widely studied in SGG task, it is also crucial to deal with the limited types and quantity of triplets in tail predicates. Accordingly, in this paper, we propose a Dual-granularity Relation Modeling (DRM) network to leverage fine-grained triplet cues besides the coarse-grained predicate ones. DRM utilizes contexts and semantics of predicate and triplet with Dual-granularity Constraints, generating compact and balanced representations from two perspectives to facilitate relation recognition. Furthermore, a Dual-granularity Knowledge Transfer (DKT) strategy is introduced to transfer variation from head predicates/triplets to tail ones, aiming to enrich the pattern diversity of tail classes to alleviate the long-tail problem. Extensive experiments demonstrate the effectiveness of our method, which establishes new state-of-the-art performance on Visual Genome, Open Image, and GQA datasets. Our code is available at \url{https://github.com/jkli1998/DRM}

Guohao Li, Hongyu Yang, Yifang Men et al.

Generating animatable and editable 3D head avatars is essential for various applications in computer vision and graphics. Traditional 3D-aware generative adversarial networks (GANs), often using implicit fields like Neural Radiance Fields (NeRF), achieve photorealistic and view-consistent 3D head synthesis. However, these methods face limitations in deformation flexibility and editability, hindering the creation of lifelike and easily modifiable 3D heads. We propose a novel approach that enhances the editability and animation control of 3D head avatars by incorporating 3D Gaussian Splatting (3DGS) as an explicit 3D representation. This method enables easier illumination control and improved editability. Central to our approach is the Editable Gaussian Head (EG-Head) model, which combines a 3D Morphable Model (3DMM) with texture maps, allowing precise expression control and flexible texture editing for accurate animation while preserving identity. To capture complex non-facial geometries like hair, we use an auxiliary set of 3DGS and tri-plane features. Extensive experiments demonstrate that our approach delivers high-quality 3D-aware synthesis with state-of-the-art controllability. Our code and models are available at https://github.com/liguohao96/EGG3D.

Xiaoxu Xu, Hao Li, Jinhui Ye et al.

Predictive foresight is important to intelligent embodied agents. Since the motor execution of a robot is intrinsically constrained by its visual perception of environmental geometry, effectively anticipating the future requires capturing this tightly coupled visuomotor interplay. While recent vision-language-action models attempt to incorporate future guidance, they struggle with this joint modeling. Existing explicit methods divert capacity to task-irrelevant visual details, whereas implicit methods relying on sparse frame pairs disrupt temporal continuity. By heavily relying on visual reconstruction, these methods become visually dominated, entangling static scene context with dynamic action intent. We argue that effective joint visuomotor predictive modeling requires both temporal continuity and visually-conditioned supervision decoupling. To this end, we propose FutureVLA, featuring a novel Joint Visuomotor Predictive Architecture. FutureVLA is designed to extract joint visuomotor embeddings by first decoupling visual and motor information, and then jointly encoding generalized physical priors. Specifically, in the pretraining stage, we leverage heterogeneous manipulation datasets and introduce a Joint Visuomotor Gating mechanism to structurally separate visual state preservation from temporal action modeling. It allows the motor stream to focus on continuous physical dynamics while explicitly querying visual tokens for environmental constraints, yielding highly generalizable joint visuomotor embeddings. Subsequently, in the post-training stage, we employ a latent embeddings alignment strategy, enabling diverse downstream VLA models to internalize these temporal priors without modifying their inference architectures. Extensive experiments demonstrate that FutureVLA consistently improves VLA frameworks.

Xun Huang, Shijia Zhao, Yunxiang Wang et al.

Embodied navigation is a fundamental capability for robotic agents operating. Real-world deployment requires open vocabulary generalization and low training overhead, motivating zero-shot methods rather than task-specific RL training. However, existing zero-shot methods that build explicit 3D scene graphs often compress rich visual observations into text-only relations, leading to high construction cost, irreversible loss of visual evidence, and constrained vocabularies. To address these limitations, we introduce the Multi-modal 3D Scene Graph (M3DSG), which preserves visual cues by replacing textual relation

Weixin Li, Yuhao Wu, Yang Liu et al.

In real recommendation scenarios, users often have different types of behaviors, such as clicking and buying. Existing research methods show that it is possible to capture the heterogeneous interests of users through different types of behaviors. However, most multi-behavior approaches have limitations in learning the relationship between different behaviors. In this paper, we propose a novel multilayer perceptron (MLP)-based heterogeneous sequential recommendation method, namely behavior-aware multilayer perceptron (BMLP). Specifically, it has two main modules, including a heterogeneous interest perception (HIP) module, which models behaviors at multiple granularities through behavior types and transition relationships, and a purchase intent perception (PIP) module, which adaptively fuses subsequences of auxiliary behaviors to capture users' purchase intent. Compared with mainstream sequence models, MLP is competitive in terms of accuracy and has unique advantages in simplicity and efficiency. Extensive experiments show that BMLP achieves significant improvement over state-of-the-art algorithms on four public datasets. In addition, its pure MLP architecture leads to a linear time complexity.

Yin Tang, Jiankai Li, Hongyu Yang et al.

In an era where social media platforms abound, individuals frequently share images that offer insights into their intents and interests, impacting individual life quality and societal stability. Traditional computer vision tasks, such as object detection and semantic segmentation, focus on concrete visual representations, while intent recognition relies more on implicit visual clues. This poses challenges due to the wide variation and subjectivity of such clues, compounded by the problem of intra-class variety in conveying abstract concepts, e.g. "enjoy life". Existing methods seek to solve the problem by manually designing representative features or building prototypes for each class from global features. However, these methods still struggle to deal with the large visual diversity of each intent category. In this paper, we introduce a novel approach named Multi-grained Compositional visual Clue Learning (MCCL) to address these challenges for image intent recognition. Our method leverages the systematic compositionality of human cognition by breaking down intent recognition into visual clue composition and integrating multi-grained features. We adopt class-specific prototypes to alleviate data imbalance. We treat intent recognition as a multi-label classification problem, using a graph convolutional network to infuse prior knowledge through label embedding correlations. Demonstrated by a state-of-the-art performance on the Intentonomy and MDID datasets, our approach advances the accuracy of existing methods while also possessing good interpretability. Our work provides an attempt for future explorations in understanding complex and miscellaneous forms of human expression.

Tiantian Cao, Xuan Dong, Chunli Peng et al.

The dual camera system of wide-angle ($\bf{W}$) and telephoto ($\bf{T}$) cameras has been widely adopted by popular phones. In the overlap region, fusing the $\bf{W}$ and $\bf{T}$ images can generate a higher quality image. Related works perform pixel-level motion alignment or high-dimensional feature alignment of the $\bf{T}$ image to the view of the $\bf{W}$ image and then perform image/feature fusion, but the enhancement in occlusion area is ill-posed and can hardly utilize data from $\bf{T}$ images. Our insight is to minimize the occlusion area and thus maximize the use of pixels from $\bf{T}$ images. Instead of insisting on placing the output in the $\bf{W}$ view, we propose a view transition method to transform both $\bf{W}$ and $\bf{T}$ images into a mixed view and then blend them into the output. The transformation ratio is kept small and not apparent to users, and the center area of the output, which has accumulated a sufficient amount of transformation, can directly use the contents from the T view to minimize occlusions. Experimental results show that, in comparison with the SOTA methods, occlusion area is largely reduced by our method and thus more pixels of the $\bf{T}$ image can be used for improving the quality of the output image.

Zixing Lei, Sheng Yin, Yichen Xiong et al.

Embodied decision-making enables agents to translate high-level goals into executable actions through continuous interactions within the physical world, forming a cornerstone of general-purpose embodied intelligence. Large language models (LLMs), with their general decision-making capabilities, offer a promising path to realize this potential; however, LLMs trained solely on language lack exposure to physical environments, limiting their true embodied understanding. To bridge this gap, we propose the concept of a training ground: a comprehensive infrastructure that provides task and scene simulation, embodied interaction, and feedback signals, offering a one-stop solution for LLM acquire genuine embodied decision-making skills. In this work, we present EmboMatrix, the first training ground of its kind, providing massive and diverse tasks with efficient simulation and precise rewards. EmboMatrix incorporates a series of novel techniques: a multi-agent data engine for large-scale task and scene generation, a distributed heterogeneous-hardware system for scalable simulation, and a multi-level reward architecture for precise supervision. Leveraging EmboMatrix, we cultivate EmboBrain, an LLM whose embodied decision-making abilities emerge from extensive embodied interactions. Experiments show that EmboBrain-7B surpasses the 671B DeepSeek-R1 baseline by 9.5\% on two challenging embodied decision-making benchmarks, demonstrating the power of interactive, environment-grounded learning for building truly intelligent embodied agents.

Zixing Lei, Zibo Zhou, Sheng Yin et al.

Humanoid whole-body control (WBC) policies trained in simulation often suffer from the sim-to-real gap, which fundamentally arises from simulator inductive bias, the inherent assumptions and limitations of any single simulator. These biases lead to nontrivial discrepancies both across simulators and between simulation and the real world. To mitigate the effect of simulator inductive bias, the key idea is to train policies jointly across multiple simulators, encouraging the learned controller to capture dynamics that generalize beyond any single simulator's assumptions. We thus introduce PolySim, a WBC training platform that integrates multiple heterogeneous simulators. PolySim can launch parallel environments from different engines simultaneously within a single training run, thereby realizing dynamics-level domain randomization. Theoretically, we show that PolySim yields a tighter upper bound on simulator inductive bias than single-simulator training. In experiments, PolySim substantially reduces motion-tracking error in sim-to-sim evaluations; for example, on MuJoCo, it improves execution success by 52.8 over an IsaacSim baseline. PolySim further enables zero-shot deployment on a real Unitree G1 without additional fine-tuning, showing effective transfer from simulation to the real world. We will release the PolySim code upon acceptance of this work.

Xuan Dong, Xiangyuan Sun, Xia Wang et al.

Moire patterns, unwanted color artifacts in images and videos, arise from the interference between spatially high-frequency scene contents and the spatial discrete sampling of digital cameras. Existing demoireing methods primarily rely on single-camera image/video processing, which faces two critical challenges: 1) distinguishing moire patterns from visually similar real textures, and 2) preserving tonal consistency and temporal coherence while removing moire artifacts. To address these issues, we propose a dual-camera framework that captures synchronized videos of the same scene: one in focus (retaining high-quality textures but may exhibit moire patterns) and one defocused (with significantly reduced moire patterns but blurred textures). We use the defocused video to help distinguish moire patterns from real texture, so as to guide the demoireing of the focused video. We propose a frame-wise demoireing pipeline, which begins with an optical flow based alignment step to address any discrepancies in displacement and occlusion between the focused and defocused frames. Then, we leverage the aligned defocused frame to guide the demoireing of the focused frame using a multi-scale CNN and a multi-dimensional training loss. To maintain tonal and temporal consistency, our final step involves a joint bilateral filter to leverage the demoireing result from the CNN as the guide to filter the input focused frame to obtain the final output. Experimental results demonstrate that our proposed framework largely outperforms state-of-the-art image and video demoireing methods.

Xinlong Cheng, Tiantian Cao, Guoan Cheng et al.

In this work, we address the limitations of denoising diffusion models (DDMs) in image restoration tasks, particularly the shape and color distortions that can compromise image quality. While DDMs have demonstrated a promising performance in many applications such as text-to-image synthesis, their effectiveness in image restoration is often hindered by shape and color distortions. We observe that these issues arise from inconsistencies between the training and testing data used by DDMs. Based on our observation, we propose a novel training method, named data-consistent training, which allows the DDMs to access images with accumulated errors during training, thereby ensuring the model to learn to correct these errors. Experimental results show that, across five image restoration tasks, our method has significant improvements over state-of-the-art methods while effectively minimizing distortions and preserving image fidelity.

Chunli Peng, Xuan Dong, Tiantian Cao et al.

The fusion of images from dual camera systems featuring a wide-angle and a telephoto camera has become a hotspot problem recently. By integrating simultaneously captured wide-angle and telephoto images from these systems, the resulting fused image achieves a wide field of view (FOV) coupled with high-definition quality. Existing approaches are mostly deep learning methods, and predominantly rely on supervised learning, where the training dataset plays a pivotal role. However, current datasets typically adopt a data synthesis approach generate input pairs of wide-angle and telephoto images alongside ground-truth images. Notably, the wide-angle inputs are synthesized rather than captured using real wide-angle cameras, and the ground-truth image is captured by wide-angle camera whose quality is substantially lower than that of input telephoto images captured by telephoto cameras. To address these limitations, we introduce a novel hardware setup utilizing a beam splitter to simultaneously capture three images, i.e. input pairs and ground-truth images, from two authentic cellphones equipped with wide-angle and telephoto dual cameras. Specifically, the wide-angle and telephoto images captured by cellphone 2 serve as the input pair, while the telephoto image captured by cellphone 1, which is calibrated to match the optical path of the wide-angle image from cellphone 2, serves as the ground-truth image, maintaining quality on par with the input telephoto image. Experiments validate the efficacy of our newly introduced dataset, named ReWiTe, significantly enhances the performance of various existing methods for real-world wide-angle and telephoto dual image fusion tasks.

Weixin Li, Jungseock Joo, Hang Qi et al.

In this paper, we aim to develop a method for automatically detecting and tracking topics in broadcast news. We present a hierarchical And-Or graph (AOG) to jointly represent the latent structure of both texts and visuals. The AOG embeds a context sensitive grammar that can describe the hierarchical composition of news topics by semantic elements about people involved, related places and what happened, and model contextual relationships between elements in the hierarchy. We detect news topics through a cluster sampling process which groups stories about closely related events. Swendsen-Wang Cuts (SWC), an effective cluster sampling algorithm, is adopted for traversing the solution space and obtaining optimal clustering solutions by maximizing a Bayesian posterior probability. Topics are tracked to deal with the continuously updated news streams. We generate topic trajectories to show how topics emerge, evolve and disappear over time. The experimental results show that our method can explicitly describe the textual and visual data in news videos and produce meaningful topic trajectories. Our method achieves superior performance compared to state-of-the-art methods on both a public dataset Reuters-21578 and a self-collected dataset named UCLA Broadcast News Dataset.

Xuan Dong, Yu Zhu, Weixin Li et al.

We study the problem of evaluating super resolution methods. Traditional evaluation methods usually judge the quality of super resolved images based on a single measure of their difference with the original high resolution images. In this paper, we proposed to use both fidelity (the difference with original images) and naturalness (human visual perception of super resolved images) for evaluation. For fidelity evaluation, a new metric is proposed to solve the bias problem of traditional evaluation. For naturalness evaluation, we let humans label preference of super resolution results using pair-wise comparison, and test the correlation between human labeling results and image quality assessment metrics' outputs. Experimental results show that our fidelity-naturalness method is better than the traditional evaluation method for super resolution methods, which could help future research on single-image super resolution.

Xuan Dong, Boyan Bonev, Weixin Li et al.

Base-detail separation is a fundamental computer vision problem consisting of modeling a smooth base layer with the coarse structures, and a detail layer containing the texture-like structures. One of the challenges of estimating the base is to preserve sharp boundaries between objects or parts to avoid halo artifacts. Many methods have been proposed to address this problem, but there is no ground-truth dataset of real images for quantitative evaluation. We proposed a procedure to construct such a dataset, and provide two datasets: Pascal Base-Detail and Fashionista Base-Detail, containing 1000 and 250 images, respectively. Our assumption is that the base is piecewise smooth and we label the appearance of each piece by a polynomial model. The pieces are objects and parts of objects, obtained from human annotations. Finally, we proposed a way to evaluate methods with our base-detail ground-truth and we compared the performances of seven state-of-the-art algorithms.