CVPR 2025 Papers

Class-Incremental Learning (CIL) enables models to continuously learn new classes while mitigating catastrophic forgetting. Recently, Pre-Trained Models (PTMs) have greatly enhanced CIL performance, even when fine-tuning is limited to the first task. This advantage is particularly beneficial for CIL methods that freeze the feature extractor after first-task fine-tuning, such as analytic learning-based approaches using a least squares solution-based classification head to acquire knowledge recursively. In this work, we revisit the analytical learning approach combined with PTMs and identify its limitations in adapting to new classes, leading to sub-optimal performance. To address this, we propose the Momentum-based Analytical Learning (MoAL) approach. MoAL achieves robust knowledge memorization via an analytical classification head and improves adaptivity to new classes through momentum-based adapter weight interpolation, leading to forgetting outdated knowledge. Importantly, we introduce a knowledge rumination mechanism that leverages refined adaptivity, allowing the model to revisit and reinforce old knowledge, thereby improving performance on old classes. MoAL facilitates the acquisition of new knowledge and consolidates old knowledge, achieving a win-win outcome between plasticity and stability. Extensive experiments on various incremental settings show MoAL's state-of-the-art performance

Previous successful approaches to missing modality completion rely on carefully designed fusion techniques and extensive pre-training on complete data, which can limit their generalizability in out-of-domain (OOD) scenarios. In this study, we pose a new challenge: can we develop a missing modality completion model that is both resource-efficient and robust to OOD generalization? To address this, we present a training-free framework for missing modality completion that leverages large multimodal models (LMMs). Our approach, termed the "Knowledge Bridger", is modality-agnostic and integrates generation and ranking of missing modalities. By defining domain-specific priors, our method automatically extracts structured information from available modalities to construct knowledge graphs. These extracted graphs connect the missing modality generation and ranking modules through the LMM, resulting in high-quality imputations of missing modalities. Experimental results across both general and medical domains show that our approach consistently outperforms competing methods, including in OOD generalization. Additionally, our knowledge-driven generation and ranking techniques demonstrate superiority over variants that directly employ LMMs for generation and ranking, offering insights that may be valuable for applications in other domains.

Existing progress in object keypoint estimation primarily benefits from the conventional supervised learning paradigm based on numerous data labeled with pre-defined keypoints. However, these well-trained models can hardly detect the undefined new keypoints in test time, which largely hinders their feasibility for diverse downstream tasks. To handle this, various solutions are explored but still suffer from either limited generalizability or transferability. Therefore, in this paper, we explore a novel keypoint learning paradigm in that we only annotate new keypoints in the new data and incrementally train the model, without retaining any old data, called Incremental object Keypoint Learning (IKL). A two-stage learning scheme as a novel baseline tailored to IKL is developed. In the first Knowledge Association stage, given the data labeled with only new keypoints, an auxiliary KA-Net is trained to automatically associate the old keypoints to these new ones based on their spatial and intrinsic anatomical relations. In the second Mutual Promotion stage, based on a keypoint-oriented spatial distillation loss, we jointly leverage the auxiliary KA-Net and the old model for knowledge consolidation to mutually promote the estimation of all old and new keypoints. Owing to the investigation of the correlations between new and old keypoints, our proposed method can not just effectively mitigate the catastrophic forgetting of old keypoints, but may even further improve the estimation of the old ones and achieve a positive transfer beyond anti-forgetting. Such an observation has been solidly verified by extensive experiments on different keypoint datasets, where our method exhibits superiority in alleviating the forgetting issue and boosting performance while enjoying labeling efficiency even under the low-shot data regime.

Learning from demonstration is a powerful method for robotic skill acquisition. However, the significant expense of collecting such action-labeled robot data presents a major bottleneck. Video data, a rich data source encompassing diverse behavioral and physical knowledge, emerges as a promising alternative. In this paper, we present GraphMimic, a novel paradigm that leverages video data via graph-to-graphs generative modeling, which pre-trains models to generate future graphs conditioned on the graph within a video frame. Specifically, GraphMimic abstracts video frames into object and visual action vertices, and constructs graphs for state representations. The graph generative modeling network then effectively models internal structures and spatial relationships within the constructed graphs, aiming to generate future graphs. The generated graphs serve as conditions for the control policy, mapping to robot actions. Our concise approach captures important spatial relations and enhances future graph generation accuracy, enabling the acquisition of robust policies from limited action-labeled data. Furthermore, the transferable graph representations facilitate the effective learning of manipulation skills from cross-embodiment videos. Our experiments exhibit that GraphMimic achieves superior performance using merely 20% action-labeled data. Moreover, our method outperforms the state-of-the-art method by over 17% and 23% in simulation and real-world experiments, and delivers improvements of over 33% in cross-embodiment transfer experiments.

Data heterogeneity in federated learning, characterized by a significant misalignment between local and global distributions, leads to divergent local optimization directions and hinders global model training. Existing studies mainly focus on optimizing local updates or global aggregation, but these indirect approaches demonstrate instability when handling highly heterogeneous data distributions, especially in scenarios where label skew and domain skew coexist. To address this, we propose a geometry-guided data generation method that centers on simulating the global embedding distribution locally. We first introduce the concept of the geometric shape of an embedding distribution and then address the challenge of obtaining global geometric shapes under privacy constraints. Subsequently, we propose GGEUR, which leverages global geometric shapes to guide the generation of new samples, enabling a closer approximation to the ideal global distribution. In single-domain scenarios, we augment samples based on global geometric shapes to enhance model generalization; in multi-domain scenarios, we further employ class prototypes to simulate the global distribution across domains. Extensive experimental results demonstrate that our method significantly enhances the performance of existing approaches in handling highly heterogeneous data, including scenarios with label skew, domain skew, and their coexistence.

The rapid development of Multimodal Large Language Models (MLLMs) has enabled the integration of multiple modalities, including texts and images, within the large language model (LLM) framework.However, texts and images are usually interconnected, forming a multimodal attributed graph (MMAG).It is underexplored how MLLMs can incorporate the relational information (i.e., graph structure) and semantic information (i.e., texts and images) on such graphs for multimodal comprehension and generation.In this paper, we propose GraphGPT-o, which supports omni-multimodal understanding and creation on MMAGs.We first comprehensively study linearization variants to transform semantic and structural information as input for MLLMs.Then, we propose a hierarchical aligner that enables deep graph encoding, bridging the gap between MMAGs and MLLMs.Finally, we explore the inference choices, adapting MLLM to interleaved text and image generation in graph scenarios. Extensive experiments on three datasets from different domains demonstrate the effectiveness of our proposed method.

Weakly Supervised Semantic Segmentation (WSSS) with image-level labels aims to achieve pixel-level predictions using Class Activation Maps (CAMs). Recently, Contrastive Language-Image Pre-training (CLIP) has been introduced in WSSS. However, recent methods primarily focus on image-text alignment for CAM generation, while CLIP's potential in patch-text alignment remains unexplored. In this work, we propose ExCEL to explore CLIP's dense knowledge via a novel patch-text alignment paradigm for WSSS. Specifically, we propose Text Semantic Enrichment (TSE) and Visual Calibration (VC) modules to improve the dense alignment across both text and vision modalities. To make text embeddings semantically informative, our TSE module applies Large Language Models (LLMs) to build a dataset-wide knowledge base and enriches the text representations with an implicit attribute-hunting process. To mine fine-grained knowledge from visual features, our VC module first proposes Static Visual Calibration (SVC) to propagate fine-grained knowledge in a non-parametric manner. Then Learnable Visual Calibration (LVC) is further proposed to dynamically shift the frozen features towards distributions with diverse semantics. With these enhancements, ExCEL not only retains CLIP's training-free advantages but also significantly outperforms other state-of-the-art methods with much less training cost on PASCAL VOC and MS COCO. Code is available at https://github.com/zwyang6/ExCEL.

As concerns regarding privacy in deep learning continue to grow, individuals are increasingly apprehensive about the potential exploitation of their personal knowledge in trained models. Despite several research efforts to address this, they often fail to consider the real-world demand from users for complete knowledge erasure. Furthermore, our investigation reveals that existing methods have a risk of leaking personal knowledge through embedding features. To address these issues, we introduce a novel concept of Knowledge Deletion (KD), an advanced task that considers both concerns, and provides an appropriate metric, named Knowledge Retention score (KR), for assessing knowledge retention in feature space. To achieve this, we propose a novel training-free erasing approach named Erasing Space Concept (ESC), which restricts the important subspace for the forgetting knowledge by eliminating the relevant activations in the feature. In addition, we suggest ESC with Training (ESC-T), which uses a learnable mask to better balance the trade-off between forgetting and preserving knowledge in KD. Our extensive experiments on various datasets and models demonstrate that our proposed methods achieve the fastest and state-of-the-art performance. Notably, our methods are applicable to diverse forgetting scenarios, such as facial domain setting, demonstrating the generalizability of our methods. The code is available at https://github.com/KU-VGI/ESC.

Diffusion models have revolutionized generative tasks through high-fidelity outputs, yet flow matching (FM) offers faster inference and empirical performance gains. However, current foundation FM models are computationally prohibitive for finetuning, while diffusion models like Stable Diffusion benefit from efficient architectures and ecosystem support. This work addresses the critical challenge of efficiently transferring knowledge from pre-trained diffusion models to flow matching. We propose Diff2Flow, a novel framework that systematically bridges diffusion and FM paradigms by rescaling timesteps, aligning interpolants, and deriving FM-compatible velocity fields from diffusion predictions. This alignment enables direct and efficient FM finetuning of diffusion priors with no extra computation overhead. Our experiments demonstrate that Diff2Flow outperforms naive FM and diffusion finetuning particularly under parameter-efficient constraints, while achieving superior or competitive performance across diverse downstream tasks compared to state-of-the-art methods.

Diffusion models (DMs) have demonstrated exceptional generative capabilities across various domains, including image, video, and so on. A key factor contributing to their effectiveness is the high quantity and quality of data used during training. However, mainstream DMs now consume increasingly large amounts of data. For example, training a Stable Diffusion model requires billions of image-text pairs. This enormous data requirement poses significant challenges for training large DMs due to high data acquisition costs and storage expenses. To alleviate this data burden, we propose a novel scenario: using existing DMs as data sources to train new DMs with any architecture. We refer to this scenario as Data-Free Knowledge Distillation for Diffusion Models (DKDM), where the generative ability of DMs is transferred to new ones in a data-free manner. To tackle this challenge, we make two main contributions. First, we introduce a DKDM objective that enables the training of new DMs via distillation, without requiring access to the data. Second, we develop a dynamic iterative distillation method that efficiently extracts time-domain knowledge from existing DMs, enabling direct retrieval of training data without the need for a prolonged generative process. To the best of our knowledge, we are the first to explore this scenario. Experimental results demonstrate that our data-free approach not only achieves competitive generative performance but also, in some instances, outperforms models trained with the entire dataset.

Lifelong person re-identification (LReID) aims to match the same person using sequentially collected data. However, due to the long-term nature of lifelong learning, the inevitable changes in human clothes prevent the model from relying on unified discriminative information (e.g., clothing style) to match the same person in the streaming data, demanding differentiated cloth-irrelevant information. Unfortunately, existing LReID methods typically fail to leverage such knowledge resulting in the exacerbation of catastrophic forgetting issues. Therefore, in this paper, we focus on a challenging practical task called Cloth-Hybrid Lifelong Person Re-identification (CH-LReID), which requires matching the same person wearing different clothes using sequentially collected data. A Differentiated Knowledge Consolidation (DKC) framework is designed to unify and balance distinct knowledge across streaming data. The core idea is to adaptively balance differentiated knowledge and compatibly consolidate cloth-relevant and cloth-irrelevant information. To this end, a Differentiated Knowledge Transfer (DKT) module and a Latent Knowledge Consolidation (LKC) module are designed to adaptively discover differentiated new knowledge, while eliminating the derived domain shift of old knowledge via reconstructing the old latent feature space, respectively. Then, to further alleviate the catastrophic conflict between differentiated new and old knowledge, we further propose a Dual-level Distribution Alignment (DDA) module to align the distribution of discriminative knowledge at both the instance level and the fine-grained level. Extensive experiments on multiple benchmarks demonstrate the superiority of our method against existing methods in both CH-LReID and traditional LReID tasks. The source code of this paper is available at https://github.com/PKU-ICST-MIPL/DKC-CVPR2025.

Long-form video understanding with Large Vision Language Models is challenged by the need to analyze temporally dispersed yet spatially concentrated key moments within limited context windows. In this work, we introduce VideoMindPalace, a new framework inspired by the "Mind Palace", which organizes critical video moments into a topologically structured semantic graph. VideoMindPalace organizes key information through (i) hand-object tracking and interaction, (ii) clustered activity zones representing specific areas of recurring activities, and (iii) environment layout mapping, allowing natural language parsing by LLMs to provide grounded insights on spatio-temporal and 3D context. In addition, we propose the Video MindPalace Benchmark (VMB), to assess human-like reasoning, including spatial localization, temporal reasoning, and layout-aware sequential understanding. Evaluated on VMB and established video QA datasets, including EgoSchema, NExT-QA, IntentQA, and the Active Memories Benchmark, VideoMindPalace demonstrates notable gains in spatio-temporal coherence and human-aligned reasoning, advancing long-form video analysis capabilities in VLMs.

Does seeing always mean knowing? Large Vision-Language Models (LVLMs) integrate separately pre-trained vision and language components, often using CLIP-ViT as vision backbone. However, these models frequently encounter a core issue of "cognitive misalignment" between the vision encoder (VE) and the large language model (LLM). Specifically, the VE's representation of visual information may not fully align with LLM's cognitive framework, leading to a mismatch where visual features exceed the language model's interpretive range. To address this, we investigate how variations in VE representations influence LVLM comprehension, especially when the LLM faces VE-Unknown data--images whose ambiguous visual representations challenge the VE's interpretive precision. Accordingly, we construct a multi-granularity landmark dataset and systematically examine the impact of VE-Known and VE-Unknown data on interpretive abilities. Our results show that VE-Unknown data limits LVLM's capacity for accurate understanding, while VE-Known data, rich in distinctive features, helps reduce cognitive misalignment. Building on these insights, we propose Entity-Enhanced Cognitive Alignment (EECA), a method that employs multi-granularity supervision to generate visually enriched, well-aligned tokens that not only integrate within the embedding space but also align with the LLM's cognitive framework. This alignment markedly enhances LVLM performance in landmark recognition. Our findings underscore the challenges posed by VE-Unknown data and highlight the essential role of cognitive alignment in advancing multimodal systems.

Advancements in large Vision-Language Models have brought precise, accurate image captioning, vital for advancing multi-modal image understanding and processing. Yet these captions often carry lengthy, intertwined contexts that are difficult to parse and frequently overlook essential cues, posing a great barrier for models like GroundingDINO and SDXL, which lack the strong text encoding and syntax analysis needed to fully leverage dense captions.To address this, we propose BACON, a prompting method that breaks down VLM-generated captions into disentangled, structured elements such as objects, relationships, styles, and themes. This approach not only minimizes confusion from handling complex contexts but also allows for efficient transfer into a JSON dictionary, enabling models without linguistic processing capabilities to easily access key information.We annotated 100,000 image-caption pairs using BACON with GPT-4V and trained an LLaVA captioner on this dataset, enabling it to produce BACON-style captions without relying on costly GPT-4V resources. Evaluations of overall quality, precision, and recall--as well as user studies--demonstrate that the resulting caption model consistently outperforms other state-of-the-art VLM models in generating high-quality captions.Additionally, we show that BACON-style captions exhibit better clarity when applied to various models, enabling them to accomplish previously unattainable tasks or surpass existing SOTA solutions without training. For example, BACON-style captions help groundingDINO achieve 1.51 times higher recall scores on open-vocabulary object detection tasks compared to leading methods.

Multimodal LLMs (MLLMs) are the natural extension of large language models to handle multimodal inputs, combining text and image data. They have recently garnered attention due to their capability to address complex tasks involving both modalities. However, their effectiveness is limited to the knowledge acquired during training, which restricts their practical utility. In this work, we introduce a novel method to enhance the adaptability of MLLMs by integrating external knowledge sources. Our proposed model, Reflective LLaVA (ReflectiVA), utilizes reflective tokens to dynamically determine the need for external knowledge and predict the relevance of information retrieved from an external database. Tokens are trained following a two-stage two-model training recipe. This ultimately enables the MLLM to manage external knowledge while preserving fluency and performance on tasks where external knowledge is not needed. Through our experiments, we demonstrate the efficacy of ReflectiVA for knowledge-based visual question answering, highlighting its superior performance compared to existing methods. Source code and trained models are publicly available at https://aimagelab.github.io/ReflectiVA.

Integrating frames and events has become a widely accepted solution for various tasks in challenging scenarios. However, most multimodal methods directly convert events into image-like formats synchronized with frames and process each stream through separate two-branch backbones, making it difficult to fully exploit the spatiotemporal events while limiting inference frequency to the frame rate. To address these problems, we propose a novel asynchronous collaborative graph representation, namely ACGR, which is the first trial to explore a unified graph framework for asynchronously processing frames and events with high performance and low latency. Technically, we first construct unimodal graphs for frames and events to preserve their spatiotemporal properties and sparsity. Then, an asynchronous collaborative alignment module is designed to align and fuse frames and events into a unified graph and the ACGR is generated through graph convolutional networks. Finally, we innovatively introduce domain adaptation to enable cross-modal interactions between frames and events by aligning their feature spaces. Experimental results show that our approach outperforms state-of-the-art methods in both object detection and depth estimation tasks, while significantly reducing computational latency and achieving real-time inference up to 200 Hz. Our code can be available at https://github.com/dianzl/ACGR.

Medical Visual Question Answering (Med-VQA) is a challenging task that requires a deep understanding of both medical images and textual questions. Although recent works leveraging Medical Vision-Language Pre-training (Med-VLP) have shown strong performance on the Med-VQA task, there is still no unified solution for modality alignment, and the issue of hard negatives remains under-explored. Additionally, commonly used knowledge fusion techniques for Med-VQA may introduce irrelevant information. In this work, we propose a framework to address these challenges through three key contributions: (1) a unified solution for heterogeneous modality alignments across multiple levels, modalities, views, and stages, leveraging methods such as contrastive learning and optimal transport theory; (2) a hard negative mining method that employs soft labels for multi-modality alignments and enforces the hard negative pair discrimination; and (3) a Gated Cross-Attention Module for Med-VQA that integrates the answer vocabulary as prior knowledge and select relevant information from it. Our framework outperforms the previous state-of-the-art on widely used Med-VQA datasets like RAD-VQA, SLAKE, PathVQA and VQA-2019. The code will be publicly available.

Vision-Language Models (VLMs) bring powerful understanding and reasoning capabilities to multimodal tasks. Meanwhile, the great need for capable aritificial intelligence on mobile devices also arises, such as the AI assistant software. Some efforts try to migrate VLMs to edge devices to expand their application scope. Simplifying the model structure is a common method, but as the model shrinks, the trade-off between performance and size becomes more and more difficult. Knowledge distillation (KD) can help models improve comprehensive capabilities without increasing size or data volume. However, most of the existing large model distillation techniques only consider applications on single-modal LLMs, or only use teachers to create new data environments for students. None of these methods take into account the distillation of the most important cross-modal alignment knowledge in VLMs. We propose a method called Align-KD to guide the student model to learn the cross-modal matching that occurs at the shallow layer. The teacher also helps student learn the projection of vision token into text embedding space based on the focus of text. Under the guidance of Align-KD, the 1.7B MobileVLM V2 model can learn rich knowledge from the 7B teacher model with light design of training loss, and achieve an average score improvement of 2.0 across 6 benchmarks under two training subsets respectively. Code is available at: https://github.com/fqhank/Align-KD.

Existing domain adaptation systems can hardly be applied to real-world problems with new classes presenting at deployment time, especially regarding source-free scenarios where multiple source domains do not share the label space despite being given a few labeled target data. To address this, we consider a challenging problem: multi-source semi-supervised open-set domain adaptation and propose a learning theory via joint error, effectively tackling strong domain shift. To generalize the algorithm into source-free cases, we introdcue a computationally efficient and architecture-flexible attention-based feature generation module. Extensive experiments on various data sets demonstrate the significant improvement of our proposed algorithm over baselines.