ICCV 2025 Papers

High-dynamic scene reconstruction aims to represent static background with rigid spatial features and dynamic objects with deformed continuous spatiotemporal features. Typically, existing methods adopt unified representation model (e.g., Gaussian) to directly match the spatiotemporal features of dynamic scene from frame camera. However, this unified paradigm fails in the potential discontinuous temporal features of objects due to frame imaging and the heterogeneous spatial features between background and objects. In this work, we introduce event camera to compensate for frame camera, and propose a spatiotemporal-disentangled Gaussian splatting framework for high-dynamic scene reconstruction. As for dynamic scene, we figure out that background and objects have appearance discrepancy in frame-based spatial features and motion discrepancy in event-based temporal features, which motivates us to distinguish the spatiotemporal features between background and objects via clustering. As for dynamic object, we discover that Gaussian representations and event data share the consistent spatiotemporal characteristic, which could serve as a prior to guide the spatiotemporal disentanglement of object Gaussians. Within Gaussian splatting framework, the cumulative scene-object disentanglement can improve the spatiotemporal discrimination between background and objects to render the time-continuous dynamic scene. Extensive experiments are performed to verify the superiority of our method.

Image diffusion models have been adapted for real-world video super-resolution to tackle over-smoothing issues in GAN-based methods. However, these models struggle to maintain temporal consistency, as they are trained on static images, limiting their ability to capture temporal dynamics effectively. Integrating text-to-video (T2V) models into video super-resolution for improved temporal modeling is straightforward. However, two key challenges remain: artifacts introduced by complex degradations in real-world scenarios, and compromised fidelity due to the strong generative capacity of powerful T2V models (e.g., CogVideoX-5B). To enhance the spatio-temporal quality of restored videos, we introduce STAR (Spatial-Temporal Augmentation with T2V models for Real-world video super-resolution), a novel approach that leverages T2V models for real-world video super-resolution, achieving realistic spatial details and robust temporal consistency. Specifically, we introduce a Local Information Enhancement Module (LIEM) before the global attention block to enrich local details and mitigate degradation artifacts. Moreover, we propose a Dynamic Frequency (DF) Loss to reinforce fidelity, guiding the modelto focus on different frequency components across diffusion steps. Extensive experiments demonstrate STAR outperforms state-of-the-art methods on both synthetic and real-world datasets.

Vector Quantization (VQ) has emerged as a prominent weight compression technique, showcasing substantially lower quantization errors than uniform quantization across diverse models, particularly in extreme compression scenarios. However, its efficacy during fine-tuning is limited by the constraint of the compression format, where weight vectors assigned to the same codeword are restricted to updates in the same direction. Consequently, many quantized weights are compelled to move in directions contrary to their local gradient information. To mitigate this issue, we introduce a novel VQ paradigm, Sign-Splitting VQ (SSVQ), which decouples the sign bit of weights from the codebook. Our approach involves extracting the sign bits of uncompressed weights and performing clustering and compression on all-positive weights. We then introduce latent variables for the sign bit and jointly optimize both the signs and the codebook. Additionally, we implement a progressive freezing strategy for the learnable sign to ensure training stability. Extensive experiments on various modern models and tasks demonstrate that SSVQ achieves a significantly superior compression-accuracy trade-off compared to conventional VQ. Furthermore, we validate our algorithm on a hardware accelerator, showing that SSVQ achieves a 3x speedup over the 8-bit compressed model by reducing memory access. Our code is available at https://github.com/list0830/SSVQ.

Accurate prediction of multi-agent future trajectories is crucial for autonomous driving systems to make safe and efficient decisions. Trajectory refinement has emerged as a key strategy to enhance prediction accuracy. However, existing refinement methods often overlook the topological relationships between trajectories, which are vital for improving prediction precision. Inspired by braid theory, we propose a novel trajectory refinement approach, Soft-Braid Refiner (SRefiner), guided by the soft-braid topological structure of trajectories using Soft-Braid Attention. Soft-Braid Attention captures spatio-temporal topological relationships between trajectories by considering both spatial proximity and vehicle motion states at "soft intersection points". Additionally, we extend this approach to model interactions between trajectories and lanes, further improving the prediction accuracy. SRefiner is a multi-iteration, multi-agent framework that iteratively refines trajectories, incorporating topological information to enhance interactions within traffic scenarios. SRefiner achieves significant performance improvements over four baseline methods across two datasets, establishing a new state-of-the-art in trajectory refinement.

Backdoor attacks have revealed the vulnerability of deep neural networks (DNNs), which motivates the development of secure deep learning systems. However, existing backdoor attacks often fail to bypass backdoor detection and human visual inspection, resulting in the exposure of the backdoor implanted in DNNs, which can subsequently be significantly mitigated through pruning or fine-tuning on benign data. To address this issue, in this paper, we propose a novel backdoor attack called SPD (Shallow Protecting Deep), which consists of a deep backdoor in the frequency domain and a shallow backdoor in the pixel domain, where the shallow backdoor acts as a firewall to protect the deep backdoor from being detected. Specifically, the deep backdoor in SPD samples from a specific Gaussian distribution, and encodes the sampled results into the intensity of the image's amplitude component in the frequency domain using an autoencoder, which serves as the backdoor trigger, thereby ensuring the invisibility of the backdoor attack. The shallow backdoor leverages traditional patch-based triggers, which covers all classes and attracts the defender's attention, thereby preserving the deep backdoor's resistance to existing backdoor detection techniques. Experimental results demonstrate that SPD not only can resist existing backdoor detection techniques, but also, due to the minimal disturbance caused by the backdoor trigger on benign samples, remains invisible, allowing the backdoor samples to pass through the human visual inspection. The source code of SPD is available at GitHub https://github.com/YuanShunJie1/SPD.

Panoptic Scene Graph Generation (PSG) integrates instance segmentation with relation understanding to capture pixel-level structural relationships in complex scenes. Although recent approaches leveraging pre-trained vision-language models (VLMs) have significantly improved performance in the open-vocabulary setting, they commonly ignore the inherent limitations of VLMs in spatial relation reasoning, such as difficulty in distinguishing object relative positions, which results in suboptimal relation prediction.Motivated by the denoising diffusion model's inversion process in preserving the spatial structure of input images, we propose SPADE (SPatial-Aware Denoising-nEtwork) framework---a novel approach for open-vocabulary PSG. SPADE consists of two key steps: (1) inversion-guided calibration for the UNet adaption, and (2) spatial-aware context reasoning. In the first step, we calibrate a general pre-trained teacher diffusion model into a PSG-specific denoising network with cross-attention maps derived during inversion through a lightweight LoRA-based fine-tuning strategy. In the second step, we develop a spatial-aware relation graph transformer that captures both local and long-range contextual information, facilitating the generation of high-quality relation queries. Extensive experiments on benchmark PSG and Visual Genome datasets demonstrate that SPADE outperforms state-of-the-art methods in both closed-set and open-set scenarios, particularly excelling in spatial relationship prediction. The code is available at: https://anonymous.4open.science/r/SPADE-105F.

Medical image segmentation data inherently contain uncertainty. This can stem from both imperfect image quality and variability in labeling preferences on ambiguous pixels, which depend on annotator expertise and the clinical context of the annotations. For instance, a boundary pixel might be labeled as tumor in diagnosis to avoid under-estimation of severity, but as normal tissue in radiotherapy to prevent damage to sensitive structures. As segmentation preferences vary across downstream applications, it is often desirable for an image segmentation model to offer user-adaptable predictions rather than a fixed output. While prior uncertainty-aware and interactive methods offer adaptability, they are inefficient at test time: uncertainty-aware models require users to choose from numerous similar outputs, while interactive models demand significant user input through click or box prompts to refine segmentation. To address these challenges, we propose SPA, a new Segmentation Preference Alignment framework that efficiently adapts to diverse test-time preferences with minimal human interaction. By presenting users with a select few, distinct segmentation candidates that best capture uncertainties, it reduces the user workload to reach the preferred segmentation. To accommodate user preference, we introduce a probabilistic mechanism that leverages user feedback to adapt a model's segmentation preference. The proposed framework is evaluated on several medical image segmentation tasks: color fundus images, lung lesion and kidney CT scans, MRI scans of brain and prostate. SPA shows 1) a significant reduction in user time and effort compared to existing interactive segmentation approaches, 2) strong adaptability based on human feedback, and 3) state-of-the-art image segmentation performance across different imaging modalities and semantic labels. Our code is publicly available here: https://github.com/SuperMedIntel/SPA.

Point transformers have demonstrated remarkable progress in 3D understanding through expanded receptive fields (RF), but further expanding the RF leads to dilution in group attention and decreases detailed feature extraction capability. Proxy, which serves as abstract representations for simplifying feature maps, enables global RF. However, existing proxy-based approaches face critical limitations: Global proxies incur quadratic complexity for large-scale point clouds and suffer positional ambiguity, while local proxy alternatives struggle with 1) Unreliable sampling from the geometrically diverse point cloud, 2) Inefficient proxy interaction computation, and 3) Imbalanced local-global information fusion; To address these challenges, we propose Sparse Proxy Point Transformer (SP^ 2 T) -- a local proxy-based dual-stream point transformer with three key innovations: First, for reliable sampling, spatial-wise proxy sampling with vertex-based associations enables robust sampling on geometrically diverse point clouds. Second, for efficient proxy interaction, sparse proxy attention with a table-based relative bias effectively achieves the interaction with efficient map-reduce computation. Third, for local-global information fusion, our dual-stream architecture maintains local-global balance through parallel branches. Comprehensive experiments reveal that SP^ 2 T sets state-of-the-art results with acceptable latency on indoor and outdoor 3D comprehension benchmarks, demonstrating marked improvement (+3.8% mIoU vs. SPoTr@S3DIS, +22.9% mIoU vs. PointASNL@Sem.KITTI) compared to other proxy-based point cloud methods.

Visual instruction tuning (VIT) enables multimodal large language models (MLLMs) to effectively handle a wide range of vision tasks by framing them as language-based instructions. Building on this, continual visual instruction tuning (CVIT) extends the capability of MLLMs to incrementally learn new tasks, accommodating evolving functionalities. While prior work has advanced CVIT through the development of new benchmarks and approaches to mitigate catastrophic forgetting, these efforts largely follow traditional continual learning paradigms, neglecting the unique challenges specific to CVIT. We identify a dual form of catastrophic forgetting in CVIT, where MLLMs not only forget previously learned visual understanding but also experience a decline in instruction following abilities as they acquire new tasks. To address this, we introduce the Separable Mixture of Low-Rank Adaptation (SMoLoRA) framework, which employs separable routing through two distinct modules--one for visual understanding and another for instruction following. This dual-routing design enables specialized adaptation in both domains, preventing forgetting while improving performance. Furthermore, we propose a new CVIT benchmark that goes beyond existing benchmarks by additionally evaluating a model's ability to generalize to unseen tasks and handle diverse instructions across various tasks. Extensive experiments demonstrate that SMoLoRA outperforms existing methods in mitigating dual forgetting, improving generalization to unseen tasks, and ensuring robustness in following diverse instructions.

Multi-modal object tracking has emerged as a significant research focus in computer vision due to its robustness in complex environments, such as exposure variations, blur, and occlusions. Despite existing studies integrating supplementary modal information into pre-trained RGB trackers through visual prompt mechanisms, this approach exhibits a critical limitation: it inherently prioritizes RGB information as the dominant modality, thereby underutilizing the complementary information of alternative modalities. To address this fundamental limitation, we present SMSTracker, an innovative tri-path score mask sigma fusion framework for multi-modal tracking, including three key modules. Firstly, we design a tri-path Score Mask Fusion (SMF) module to evaluate and quantify the reliability of each modality, allowing optimal exploitation of complementary features between modalities. Secondly, we introduce a pioneering Sigma Interaction (SGI) module to facilitate a sophisticated fusion of modal features across tri-branches. Furthermore, we advance a Drop Key Fine-tuning (DKF) strategy to address the inherent challenge of unequal data contribution in multi-modal learning scenarios, thereby enhancing the model's capacity for comprehensive multi-modal information processing. Finally, extensive experiments on RGB+Thermal, RGB+Depth, and RGB+Event datasets demonstrate the significant performance improvements achieved by SMSTracker over existing state-of-the-art methods. Code and model are available at https://github.com/Leezed525/SMSTracker.

Transfer-based attacks pose a significant security threat to deep neural networks (DNNs), due to their strong performance on unseen models in real-world black-box scenarios. Building on this, feature importance-based attacks further improve the transferability of adversarial examples by effectively suppressing model-specific feature patterns. However, existing methods primarily focus on single-granularity patch and single-stage training, leading to suboptimal solutions. To address these limitations, we propose a general multi-stage optimization framework based on Semantics-aware Multi-granularity Patchout, dubbed as SMP-Attack. Compared to the non-deformable/regular patch definition, we incorporate multi-granularity into the generation process of deformable/irregular patches, thereby enhancing the quality of the computed aggregate gradient. In contrast to conventional joint optimization of multi-layer losses, we introduce an effective multi-stage training strategy that systematically explores significant model-agnostic features from shallow to intermediate layers. Employing the ImageNet dataset, we conduct extensive experiments on undefended/defended CNNs and ViTs, which unequivocally demonstrate the superior performance of our proposed SMP-Attack over current state-of-the-art methods in black-box scenarios. Furthermore, we assess the compatibility of our multi-stage optimization, which supersedes single-stage training employed in existing feature-based methods, culminating in substantial performance improvement. Code is available at https://github.com/AdvML-Group/SMP-Attack.

Soccer is a globally renowned sport with significant applications in video games and VR/AR. However, generating realistic soccer motions remains challenging due to the intricate interactions between the player and the ball. In this paper, we introduce SMGDiff, a novel two-stage framework for generating real-time and user-controllable soccer motions. Our key idea is to integrate real-time character animation with a powerful diffusion-based generative model. Specifically, we first map coarse user control to intricate character trajectories. Then, we employ a transformer-based autoregressive diffusion model to generate soccer motions based on trajectory conditioning. For further physical realism, we integrate a contact guidance module during inference to refine precise ball-foot interactions.Additionally, we contribute a large-scale soccer motion dataset consisting of over 1.08 million frames of diverse soccer motions. Extensive experiments demonstrate that our SMGDiff significantly outperforms existing methods in terms of motion quality and condition alignment.

From optical sensors to microwave radars, leveraging the complementary strengths of remote sensing (RS) sensors is crucial for achieving dense spatio-temporal monitoring of our planet. In contrast, recent deep learning models, whether task-specific or foundational, are often specific to single sensors or to fixed combinations: adapting such models to different sensory inputs requires both architectural changes and re-training, limiting scalability and generalization across multiple RS sensors. On the contrary, a single model able to modulate its feature representations to accept diverse sensors as input would pave the way to agile and flexible multi-sensor RS data processing. To address this, we introduce SMARTIES, a generic and versatile foundation model lifting sensor-specific/dependent efforts and enabling scalability and generalization to diverse RS sensors: SMARTIES projects data from heterogeneous sensors into a shared spectrum-aware space, enabling the use of arbitrary combinations of bands both for training and inference. To obtain sensor-agnostic representations, we train a single, unified transformer model reconstructing masked multi-sensor data with cross-sensor token mixup. On both single- and multi-modal tasks across diverse sensors, SMARTIES outperforms previous models that rely on sensor-specific pretraining. Our code and pretrained models are available at https://gsumbul.github.io/SMARTIES.

Implicit Neural Representation (INR), leveraging a neural network to transform coordinate input into corresponding attributes, has recently driven significant advances in several vision-related domains. However, the performance of INR is heavily influenced by the choice of the nonlinear activation function used in its multilayer perceptron (MLP) architecture. To date, multiple nonlinearities have been investigated, but current INRs still face limitations in capturing high-frequency components and diverse signal types. We show that these challenges can be alleviated by introducing a novel approach in INR architecture. Specifically, we propose SL^ 2 A-INR, a hybrid network that combines a single-layer learnable activation function with an MLP that uses traditional ReLU activations. Our method performs superior across diverse tasks, including image representation, 3D shape reconstruction, and novel view synthesis. Through comprehensive experiments, SL^ 2 A-INR sets new benchmarks in accuracy, quality, and robustness for INR.

We present SKALD, a multi-shot video assembly method that constructs coherent video sequences from candidate shots with minimal reliance on text. Central to our approach is the Learned Clip Assembly (LCA) score, a learning-based metric that measures temporal and semantic relationships between shots to quantify narrative coherence. We tackle the exponential complexity of combining multiple shots with an efficient beam-search algorithm guided by the LCA score. To train our model effectively with limited human annotations, we propose two tasks for the LCA encoder: Shot Coherence Learning, which uses contrastive learning to distinguish coherent and incoherent sequences, and Feature Regression, which converts these learned representations into a real-valued coherence score. We develop two variants: a base SKALD model that relies solely on visual coherence and SKALD-text, which integrates auxiliary text information when available. Experiments on the VSPD and our curated MSV3C datasets show that SKALD achieves an improvement of up to 48.6% in IoU and a 43% speedup over the state-of-the-art methods. A user study further validates our approach, with 45% of participants favoring SKALD-assembled videos, compared to 22% preferring text-based assembly methods.

Spatial intelligence (SI) represents a cognitive ability encompassing the visualization, manipulation, and reasoning about spatial relationships, underpinning disciplines from neuroscience to robotics. We introduce SITE, a benchmark dataset towards SI Thorough Evaluation in a standardized format of multi-choice visual question-answering, designed to assess large vision-language models' spatial intelligence across diverse visual modalities (single-image, multi-image, and video) and SI factors (figural to environmental scales, spatial visualization and orientation, intrinsic and extrinsic, static and dynamic). Our approach to curating the benchmark combines a bottom-up survey of existing datasets and a top-down strategy drawing upon three classification systems in cognitive science, which prompt us to design two novel types of tasks about view-taking and dynamic scenes. Extensive experiments reveal that leading models fall behind human experts, especially in spatial orientation, a fundamental SI factor. Moreover, we demonstrate a positive correlation between a model's spatial reasoning proficiency and its performance on an embodied AI task.

Simulating stylized human-scene interactions (HSI) in physical environments is a challenging yet fascinating task. Prior works emphasize long-term execution but fall short in achieving both diverse style and physical plausibility. To tackle this challenge, we introduce a novel hierarchical framework named SIMS that seamlessly bridges high-level script-driven intent with a low-level control policy, enabling more expressive and diverse human-scene interactions. Specifically, we employ Large Language Models with Retrieval-Augmented Generation (RAG) to generate coherent and diverse long-form scripts, providing a rich foundation for motion planning. A versatile multi-condition physics-based control policy is also developed, which leverages text embeddings from the generated scripts to encode stylistic cues, simultaneously perceiving environmental geometries and accomplishing task goals. By integrating the retrieval-augmented script generation with the multi-condition controller, our approach provides a unified solution for generating stylized HSI motions. We further introduce a comprehensive planning dataset produced by RAG and a stylized motion dataset featuring diverse locomotions and interactions. Extensive experiments demonstrate SIMS's effectiveness in executing various tasks and generalizing across different scenarios, significantly outperforming previous methods.

Plug-and-play methods for solving inverse problems have continuously improved over the years by incorporating more advanced image priors.Latent diffusion models are among the most powerful priors, making them a natural choice for solving inverse problems. However, existing approaches require multiple applications of an Autoencoder to transition between pixel and latent spaces during restoration, leading to high computational costs and degraded restoration quality. In this work, we introduce a new plug-and-play paradigm that operates entirely in the latent space of diffusion models. By emulating pixel-space degradations directly in the latent space through a short learning phase, we eliminate the need for the Autoencoder during restoration, enabling faster inference and improved restoration fidelity. We validate our method across various image restoration tasks and datasets, achieving significantly higher perceptual quality than previous methods while being 2.6-10x faster in inference and 1.7-7x faster when accounting for the learning phase of the latent operator.