CVPR 2025 论文列表

Federated learning has wide applications in the medical field. It enables knowledge sharing among different healthcare institutes while protecting patients' privacy. However, existing federated learning systems are typically centralized, requiring clients to upload client-specific knowledge to a central server for aggregation. This centralized approach would integrate the knowledge from each client into a centralized server, and the knowledge would be already undermined during the centralized integration before it reaches back to each client. Besides, the centralized approach also creates a dependency on the central server, which may affect training stability if the server malfunctions or connections are unstable. To address these issues, we propose a decentralized federated learning framework named dFLMoE. In our framework, clients directly exchange lightweight head models with each other. After exchanging, each client treats both local and received head models as individual experts, and utilizes a client-specific Mixture of Experts (MoE) approach to make collective decisions. This design not only reduces the knowledge damage with client-specific aggregations but also removes the dependency on the central server to enhance the robustness of the framework. We validate our framework on multiple medical tasks, demonstrating that our method evidently outperforms state-of-the-art approaches under both model homogeneity and heterogeneity settings.

We present V2Dial - a novel expert-based model specifically geared towards simultaneously handling image and video input data for multimodal conversational tasks. Current multimodal models primarily focus on simpler tasks (e.g., VQA, VideoQA, video-text retrieval) and often neglect the more challenging conversational counterparts, such as video and visual/image dialog. Moreover, works on both conversational tasks evolved separately from each other despite their apparent similarities limiting their applicability potential. To this end, we propose to unify both tasks using a single model that for the first time jointly learns the spatial and temporal features of images and videos by routing them through dedicated experts and aligns them using matching and contrastive learning techniques. Furthermore, we systemically study the domain shift between the two tasks by investigating whether and to what extent these seemingly related tasks can mutually benefit from their respective training data. Extensive evaluations on the widely used video and visual dialog datasets of AVSD and VisDial show that our model achieves new state-of-the-art results across four benchmarks both in zero-shot and fine-tuning settings.

Continual learning (CL) aims to continually accumulate knowledge from a non-stationary data stream without catastrophic forgetting of learned knowledge, requiring a balance between stability and adaptability. Relying on the generalizable representation in pre-trained models (PTMs), PTM-based CL methods perform effective continual adaptation on downstream tasks by adding learnable adapters or prompts upon the frozen PTMs. However, many existing PTM-based CL methods use restricted adaptation on a fixed set of these modules to avoid forgetting, suffering from limited CL ability. Periodically adding task-specific modules results in linear model growth rate and impaired knowledge reuse. We propose Self-Expansion of pre-trained models with Modularized Adaptation (SEMA), a novel approach to enhance the control of stability-plasticity balance in PTM-based CL. SEMA automatically decides to reuse or add adapter modules on demand in CL, depending on whether significant distribution shift that cannot be handled is detected at different representation levels. We design modular adapter consisting of a functional adapter and a representation descriptor. The representation descriptors are trained as a distribution shift indicator and used to trigger self-expansion signals. For better composing the adapters, an expandable weighting router is learned jointly for mixture of adapter outputs. SEMA enables better knowledge reuse and sub-linear expansion rate. Extensive experiments demonstrate the effectiveness of the proposed self-expansion method, achieving state-of-the-art performance compared to PTM-based CL methods without memory rehearsal. Code is available at https://github.com/huiyiwang01/SEMA-CL.

Most state-of-the-art trackers adopt one-stream paradigm, using a single Vision Transformer for joint feature extraction and relation modeling of template and search region images. However, relation modeling between different image patches exhibits significant variations. For instance, background regions dominated by target-irrelevant information require reduced attention allocation, while foreground, particularly boundary areas, need to be be emphasized. A single model may not effectively handle all kinds of relation modeling simultaneously. In this paper, we propose a novel tracker called SPMTrack based on mixture-of-experts tailored for visual tracking task (TMoE), combining the capability of multiple experts to handle diverse relation modeling more flexibly. Benefiting from TMoE, we extend relation modeling from image pairs to spatio-temporal context, further improving tracking accuracy with minimal increase in model parameters. Moreover, we employ TMoE as a parameter-efficient fine-tuning method, substantially reducing trainable parameters, which enables us to train SPMTrack of varying scales efficiently and preserve the generalization ability of pretrained models to achieve superior performance. We conduct experiments on seven datasets, and experimental results demonstrate that our method significantly outperforms current state-of-the-art trackers. The source code is available at https://github.com/WenRuiCai/SPMTrack.

Audio-Visual Question Answering (AVQA) requires not only question-based multimodal reasoning but also precise temporal grounding to capture subtle dynamics for accurate prediction. However, existing methods mainly use question information implicitly, limiting focus on question-specific details. Furthermore, most studies rely on uniform frame sampling, which can miss key question-relevant frames. Although recent Top-K frame selection methods aim to address this, their discrete nature still overlooks fine-grained temporal details. This paper proposes QA-TIGER, a novel framework that explicitly incorporates question information and models continuous temporal dynamics. Our key idea is to use Gaussian-based modeling to adaptively focus on both consecutive and non-consecutive frames based on the question, while explicitly injecting question information and applying progressive refinement. We leverage a Mixture of Experts (MoE) to flexibly implement multiple Gaussian models, activating temporal experts specifically tailored to the question. Extensive experiments on multiple AVQA benchmarks show that QA-TIGER consistently achieves state-of-the-art performance.

Object detection in event streams has emerged as a cutting-edge research area, demonstrating superior performance in low-light conditions, scenarios with motion blur, and rapid movements. Current detectors leverage spiking neural networks, Transformers, or convolutional neural networks as their core architectures, each with its own set of limitations including restricted performance, high computational overhead, or limited local receptive fields. This paper introduces a novel MoE (Mixture of Experts) heat conduction-based object detection algorithm that strikingly balances accuracy and computational efficiency. Initially, we employ a stem network for event data embedding, followed by processing through our innovative MoE-HCO blocks. Each block integrates various expert modules to mimic heat conduction within event streams. Subsequently, an IoU-based query selection module is utilized for efficient token extraction, which is then channeled into a detection head for the final object detection process. Furthermore, we are pleased to introduce EvDET200K, a novel benchmark dataset for event-based object detection. Captured with a high-definition Prophesee EVK4-HD event camera, this dataset encompasses 10 distinct categories, 200,000 bounding boxes, and 10,054 samples, each spanning 2 to 5 seconds. We also provide comprehensive results from over 15 state-of-the-art detectors, offering a solid foundation for future research and comparison.

Visual encoders are fundamental components in vision-language models (VLMs), each showcasing unique strengths derived from various pre-trained visual foundation models. To leverage the various capabilities of these encoders, recent studies incorporate multiple encoders within a single VLM, leading to a considerable increase in computational cost. In this paper, we present Mixture-of-Visual-Encoder Knowledge Distillation (MoVE-KD), a novel framework that distills the unique proficiencies of multiple vision encoders into a single, efficient encoder model. Specifically, to mitigate conflicts and retain the unique characteristics of each teacher encoder, we employ low-rank adaptation (LoRA) and mixture-of-experts (MoEs) to selectively activate specialized knowledge based on input features, enhancing both adaptability and efficiency. To regularize the KD process and enhance performance, we propose an attention-based distillation strategy that adaptively weighs the different encoders and emphasizes valuable visual tokens, reducing the burden of replicating comprehensive but distinct features from multiple teachers. Comprehensive experiments on popular VLMs, such as LLaVA and LLaVA-NeXT, validate the effectiveness of our method. Our code is available at:https://github.com/hey-cjj/MoVE-KD.

The generation of talking avatars has achieved significant advancements in precise audio synchronization. However, crafting lifelike talking head videos requires capturing a broad spectrum of emotions and subtle facial expressions. Current methods face fundamental challenges: a) the absence of frameworks for modeling single basic emotional expressions, which restricts the generation of complex emotions such as compound emotions; b) the lack of comprehensive datasets rich in human emotional expressions, which limits the potential of models. To address these challenges, we propose the following innovations: 1) the Mixture of Emotion Experts (MoEE) model, which decouples six fundamental emotions to enable the precise synthesis of both singular and compound emotional states; 2) the DH-FaceEmoVid-150 dataset, specifically curated to include six prevalent human emotional expressions as well as four types of compound emotions, thereby expanding the training potential of emotion-driven models; 3) an emotion-to-latents module that leverages multimodal inputs, aligning diverse control signals--such as audio, text, and labels--to enhance audio-driven emotion control. Through extensive quantitative and qualitative evaluations, we demonstrate that the MoEE framework, in conjunction with the DH-FaceEmoVid-150 dataset, excels in generating complex emotional expressions and nuanced facial details, setting a new benchmark in the field. These datasets will be publicly released.

Existing technique on domain adaptive person search commonly utilizes the unified framework for jointly localizing and identifying the person across domains. This framework, however, inevitably results in the gradient conflict problem, particularly in cross-domain scenarios with contradictory objectives, as the unified framework employs shared parameters to simultaneously address person detection and re-identification tasks across the domains. To overcome this, we present a novel mixture of submodules framework, dubbed MoS, that dynamically modulates the combination of submodules depending on the specific task to perform person detection and re-identification, separately. We further design the mixtures of submodules that vary depending on the domain, enabling domain-specific knowledge transfer. Especially, we decompose the main model into several submodules and employ diverse mixtures of submodules that vary depending on the tasks and domains through the conditional routing policy. In addition, we also present counterpart domain sample generation that synthesizes the augmented sample and uses them to learn domain invariant representation for person re-identification through the contrastive domain alignment. We conduct experiments to demonstrate the effectiveness of our MoS over the existing domain adaptive person search method and provide ablation studies.

Model editing aims to correct inaccurate knowledge, update outdated information, and incorporate new data into Large Language Models (LLMs) without the need for retraining. This task poses challenges in lifelong scenarios where edits must be continuously applied for real-world applications. While some editors demonstrate strong robustness for lifelong editing in pure LLMs, Vision LLMs (VLLMs), which incorporate an additional vision modality, are not directly adaptable to existing LLM editors. In this paper, we propose LiveEdit, a lifelong vision language model edit to bridge the gap between lifelong LLM editing and VLLMs. We begin by training an editing expert generator to independently produce low-rank experts for each editing instance, with the goal of correcting the relevant responses of the VLLM. A hard filtering mechanism is developed to utilize visual semantic knowledge, thereby coarsely eliminating visually irrelevant experts for input queries during the inference stage of the post-edited model. Finally, to integrate visually relevant experts, we introduce a soft routing mechanism based on textual semantic relevance to achieve multi-expert fusion. For evaluation, we establish a benchmark for lifelong VLLM editing. Extensive experiments demonstrate that LiveEdit offers significant advantages in lifelong VLLM editing scenarios. Further experiments validate the rationality and effectiveness of each module design in LiveEdit.

LiDAR data pretraining offers a promising approach to leveraging large-scale, readily available datasets for enhanced data utilization. However, existing methods predominantly focus on sparse voxel representation, overlooking the complementary attributes provided by other LiDAR representations. In this work, we propose LiMoE, a framework that integrates the Mixture of Experts (MoE) paradigm into LiDAR data representation learning to synergistically combine multiple representations, such as range images, sparse voxels, and raw points. Our approach consists of three stages: i) Image-to-LiDAR Pretraining, which transfers prior knowledge from images to point clouds across different representations; ii) Contrastive Mixture Learning (CML), which uses MoE to adaptively activate relevant attributes from each representation and distills these mixed features into a unified 3D network; iii) Semantic Mixture Supervision (SMS), which combines semantic logits from multiple representations to boost downstream segmentation performance. Extensive experiments across eleven large-scale LiDAR datasets demonstrate our effectiveness and superiority. The code has been made publicly accessible.

The number of publicly available models is rapidly increasing, yet most remain undocumented. Users looking for suitable models for their tasks must first determine what each model does. Training machine learning models to infer missing documentation directly from model weights is challenging, as these weights often contain significant variation unrelated to model functionality (denoted nuisance). Here, we identify a key property of real-world models: most public models belong to a small set of Model Trees, where all models within a tree are fine-tuned from a common ancestor (e.g., a foundation model). Importantly, we find that within each tree there is less nuisance variation between models. Concretely, while learning across Model Trees requires complex architectures, even a linear classifier trained on a single model layer often works within trees. While effective, these linear classifiers are computationally expensive, especially when dealing with larger models that have many parameters. To address this, we introduce Probing Experts (ProbeX), a theoretically motivated and lightweight method. Notably, ProbeX is the first probing method specifically designed to learn from the weights of a single hidden model layer. We demonstrate the effectiveness of ProbeX by predicting the categories in a model's training dataset based only on its weights. Excitingly, ProbeX can map the weights of Stable Diffusion into a weight-language embedding space, enabling model search via text, i.e., zero-shot model classification.

Analyzing gigapixel Whole Slide Images (WSIs) is challenging due to the complex pathological tissue environment and the absence of target-driven domain knowledge. Previous methods incorporated pathological priors to mitigate this issue but relied on additional inference steps and specialized workflows, restricting scalability and the model's capacity to identify novel outcome-related factors. To address these challenges, we propose a plug-and-play Pathology-Aware Mixture-of-Experts (PAMoE) module, which based on mixture of experts to learn pathology-related knowledge and extract useful information. We train the experts to become 'specialists' in specific intratumoral tissues by learning to route each tissue to its mapped expert. In addition, to reduce the impact of irrelevant content on the model, we introduce a new routing rule that discards patches in which none of the experts express interest, which helps the model better capture the relationships between relevant patches. Through a comprehensive evaluation of PAMoE on survival task, we demonstrate that 1) Our module enhances the performance of baseline models in most cases, and 2) The sparse expert processing across different tissues enhances the learning of patch representations by addressing tissue heterogeneity.

Online continual learning (OCL) seeks to learn new tasks from data streams that appear only once, while retaining knowledge of previously learned tasks. Most existing methods rely on replay, focusing on enhancing memory retention through regularization or distillation. However, they often overlook the adaptability of the model, limiting the ability to learn generalizable and discriminative features incrementally from online training data.To address this, we introduce a plug-and-play module, S6MOD, which can be integrated into most existing methods and directly improve adaptability. Specifically, S6MOD introduces an extra branch after the backbone, where a mixture of discretization selectively adjusts parameters in a selective state space model, enriching selective scan patterns such that the model can adaptively select the most sensitive discretization method for current dynamics.We further design a class-conditional routing algorithm for dynamic, uncertainty-based adjustment and implement a contrastive discretization loss to optimize it. Extensive experiments combining our module with various models demonstrate that S6MOD significantly enhances model adaptability, leading to substantial performance gains and achieving the state-of-the-art results. The code is available at https://github.com/MyToumaKazusa/S6MOD.

Vision Transformers (ViTs) have emerged as the state-of-the-art models in various Computer Vision (CV) tasks, but their high computational and resource demands pose significant challenges. While Mixture of Experts (MoE) can make these models more efficient, they often require costly retraining or even training from scratch. Recent developments aim to reduce these computational costs by leveraging pretrained networks. These have been shown to produce sparse activation patterns in the Multi-Layer Perceptrons (MLPs) of the encoder blocks, allowing for conditional activation of only relevant subnetworks for each sample. Building on this idea, we propose a new method to construct MoE variants from pretrained models. Our approach extracts expert subnetworks from the model's MLP layers post-training in two phases. First, we cluster output activations to identify distinct activation patterns. In the second phase, we use these clusters to extract the corresponding subnetworks responsible for producing them. On ImageNet-1k recognition tasks, we demonstrate that these extracted experts can perform surprisingly well out of the box and require only minimal fine-tuning to regain 98% of the original performance, all while reducing FLOPs and model size, by up to 36% and 32% respectively.

Multimodal Emotion Recognition (MER) aims to predict human emotions by leveraging multiple modalities, such as vision, acoustics, and language. However, due to the heterogeneity of these modalities, MER faces two key challenges: modality balance dilemma and modality specialization disappearance. Existing methods often overlook the varying importance of modalities across samples in tackling the modality balance dilemma. Moreover, mainstream decoupling methods, while preserving modality-specific information, often neglect the predictive capability of unimodal data. To address these, we propose a novel model, Modality-Specific Enhanced Dynamic Emotion Experts (EMOE), consisting of: (1) Mixture of Modality Experts for dynamically adjusting modality importance based on sample features, and (2) Unimodal Distillation to retain single-modality predictive ability within fused features. EMOE enables adaptive fusion by learning a unique modality weight distribution for each sample, enhancing multimodal predictions with single-modality predictions to balance invariant and specific features in emotion recognition. Experimental results on benchmark datasets show that EMOE achieves superior or comparable performance to state-of-the-art methods. Additionally, we extend EMOE to Multimodal Intent Recognition (MIR), further demonstrating its effectiveness and versatility.

Upcycled Mixture-of-Experts (MoE) models have shown great potential in various tasks by converting the original Feed-Forward Network (FFN) layers in pre-trained dense models into MoE layers. However, these models still suffer from significant parameter inefficiency due to the introduction of multiple experts. In this work, we propose a novel DeRS (Decompose, Replace, and Synthesis) paradigm to overcome this shortcoming, which is motivated by our observations about the unique redundancy mechanisms of upcycled MoE experts. Specifically, DeRS decomposes the experts into one expert-shared base weight and multiple expert-specific delta weights, and subsequently represents these delta weights in lightweight forms. Our proposed DeRS paradigm can be applied to enhance parameter efficiency in two different scenarios, including: 1) DeRS Compression for inference stage, using sparsification or quantization to compress vanilla upcycled MoE models; and 2) DeRS Upcycling for training stage, employing lightweight sparse or low-rank matrixes to efficiently upcycle dense models into MoE models. Extensive experiments across three different tasks show that the proposed methods can achieve extreme parameter efficiency while maintaining the performance for both training and compression of upcycled MoE models.

Recent advancements in all-in-one image restoration models have revolutionized the ability to address diverse degradations through a unified framework. However, parameters tied to specific tasks often remain inactive for other tasks, making mixture-of-experts (MoE) architectures a natural extension. Despite this, MoEs often show inconsistent behavior, with some experts unexpectedly generalizing across tasks while others struggle within their intended scope. This hinders leveraging MoEs' computational benefits by bypassing irrelevant experts during inference.We attribute this undesired behavior to the uniform and rigid architecture of traditional MoEs. To address this, we introduce "complexity experts" -- flexible expert blocks with varying computational complexity and receptive fields. A key challenge is assigning tasks to each expert, as degradation complexity is unknown in advance. Thus, we execute tasks with a simple bias toward lower complexity.To our surprise, this preference effectively drives task-specific allocation, assigning tasks to experts with the appropriate complexity. Extensive experiments validate our approach, demonstrating the ability to bypass irrelevant experts during inference while maintaining superior performance. The proposed MoCE-IR model outperforms state-of-the-art methods, affirming its efficiency and practical applicability. The source code and models are publicly available at \href https://eduardzamfir.github.io/moceir/ \texttt eduardzamfir.github.io/MoCE-IR/

The goal of point cloud localization based on linguistic description is to identify a 3D position using textual description in large urban environments, which has potential applications in various fields, such as determining the location for vehicle pickup or goods delivery. Ideally, for a textual description and its corresponding 3D location, the objects around the 3D location should be fully described in the text description. However, in practical scenarios, e.g., vehicle pickup, passengers usually describe only the part of the most significant and nearby surroundings instead of the entire environment. In response to this partially relevant challenge, we propose CMMLoc, an uncertainty-aware Cauchy-Mixture-Model (CMM) based framework for text-to-point-cloud Localization. To model the uncertain semantic relations between text and point cloud, we integrate CMM constraints as a prior during the interaction between the two modalities. We further design a spatial consolidation scheme to enable adaptive aggregation of different 3D objects with varying receptive fields. To achieve precise localization, we propose a cardinal direction integration module alongside a modality pre-alignment strategy, helping capture the spatial relationships among objects and bringing the 3D objects closer to the text modality. Comprehensive experiments validate that CMMLoc outperforms existing methods, achieving state-of-the-art results on the KITTI360Pose dataset. Codes are available in this anonymous GitHub repository https://github.com/anonymous0819/CMMLoc.