CVPR 2025 Papers

3D semantic occupancy prediction has garnered attention as an important task for the robustness of vision-centric autonomous driving, which predicts fine-grained geometry and semantics of the surrounding scene. Most existing methods leverage dense grid-based scene representations, overlooking the spatial sparsity of the driving scenes, which leads to computational redundancy. Although 3D semantic Gaussian serves as an object-centric sparse alternative, most of the Gaussians still describe the empty region with low efficiency. To address this, we propose a probabilistic Gaussian superposition model which interprets each Gaussian as a probability distribution of its neighborhood being occupied and conforms to probabilistic multiplication to derive the overall geometry. Furthermore, we adopt the exact Gaussian mixture model for semantics calculation to avoid unnecessary overlapping of Gaussians. To effectively initialize Gaussians in non-empty region, we design a distribution-based initialization module which learns the pixel-aligned occupancy distribution instead of the depth of surfaces. We conduct extensive experiments on nuScenes and KITTI-360 datasets and our GaussianFormer-2 achieves state-of-the-art performance with high efficiency.

Masked Image Modeling (MIM) has emerged as a powerful self-supervised learning paradigm for visual representation learning, enabling models to acquire rich visual representations by predicting masked portions of images from their visible regions. While this approach has shown promising results, we hypothesize that its effectiveness may be limited by optimization challenges during early training stages, where models are expected to learn complex image distributions from partial observations before developing basic visual processing capabilities. To address this limitation, we propose a prototype-driven curriculum leagrning framework that structures the learning process to progress from prototypical examples to more complex variations in the dataset. Our approach introduces a temperature-based annealing scheme that gradually expands the training distribution, enabling more stable and efficient learning trajectories. Through extensive experiments on ImageNet-1K, we demonstrate that our curriculum learning strategy significantly improves both training efficiency and representation quality while requiring substantially fewer training epochs compared to standard Masked Auto-Encoding. Our findings suggest that carefully controlling the order of training examples plays a crucial role in self-supervised visual learning, providing a practical solution to the early-stage optimization challenges in MIM.

Model Inversion Attacks (MIAs) aim to reconstruct private training data from models, leading to privacy leakage, particularly in facial recognition systems. Although many studies have enhanced the effectiveness of white-box MIAs, less attention has been paid to improving efficiency and utility under limited attacker capabilities. Existing black-box MIAs necessitate an impractical number of queries, incurring significant overhead. Therefore, we analyze the limitations of existing MIAs and introduce Surrogate Model-based Inversion with Long-tailed Enhancement (SMILE), a high-resolution oriented and query-efficient MIA for the black-box setting. We begin by analyzing the initialization of MIAs from a data distribution perspective and propose a long-tailed surrogate training method to obtain high-quality initial points. We then enhance the attack's effectiveness by employing the gradient-free black-box optimization algorithm selected by NGOpt. Our experiments show that SMILE outperforms existing state-of-the-art black-box MIAs while requiring only about 5% of the query overhead.

Video-language models have shown promise for addressing a range of multimodal tasks for video understanding, such as video question-answering. However, the inherent computational challenges of processing long video data and increasing model sizes have led to standard approaches that are limited by the number of frames they can process. In this work, we propose the Flexible Frame Selector (FFS), a learnable policy model with a new flexible selection operation, that helps alleviate input context restrictions by enabling video-language models to focus on the most informative frames for the downstream multimodal task, without adding undue processing cost. Our method differentiates from prior work due to its learnability, efficiency, and flexibility. We verify the efficacy of our method on standard video-question answering and reasoning benchmarks, and observe that our model can improve base video-language model accuracy while reducing the number of downstream processed frames.

In the realm of high-frequency data streams, achieving real-time learning within varying memory constraints is paramount. This paper presents Ferret, a comprehensive framework designed to enhance online accuracy of Online Continual Learning (OCL) algorithms while dynamically adapting to varying memory budgets.Ferret employs a fine-grained pipeline parallelism strategy combined with an iterative gradient compensation algorithm, ensuring seamless handling of high-frequency data with minimal latency, and effectively counteracting the challenge of stale gradients in parallel training. To adapt to varying memory budgets, its automated model partitioning and pipeline planning optimizes performance regardless of memory limitations. Extensive experiments across 20 benchmarks and 5 integrated OCL algorithms show Ferret's remarkable efficiency, achieving up to 3.7x lower memory overhead to reach the same online accuracy compared to competing methods.Furthermore, Ferret consistently outperforms these methods across diverse memory budgets, underscoring its superior adaptability. These findings position Ferret as a premier solution for efficient and adaptive OCL framework in real-time environments.

Parameter-efficient tuning (PET) aims to transfer pre-trained foundation models to downstream tasks by learning a small number of parameters. Compared to traditional fine-tuning, which updates the entire model, PET significantly reduces storage and transfer costs for each task regardless of exponentially increasing pre-trained model capacity. However, most PET methods inherit the inference latency of their large backbone models and often introduce additional computational overhead due to additional modules (e.g. adapters), limiting their practicality for compute-intensive applications. In this paper, we propose Faster Parameter-Efficient Tuning (FPET), a novel approach that enhances inference speed and training efficiency while maintaining high storage efficiency. Specifically, we introduce a plug-and-play token redundancy reduction module delicately designed for PET. This module refines tokens from the self-attention layer using an adapter to learn the accurate similarity between tokens and cuts off the tokens through a fully-differentiable token merging strategy, which uses a straight-through estimator for optimal token reduction. Experimental results prove that our FPET achieves faster inference and higher memory efficiency than the pre-trained backbone while keeping competitive performance on par with state-of-the-art PET methods.

Language-image pre-training faces significant challenges due to limited data in specific formats and the constrained capacities of text encoders. While prevailing methods attempt to address these issues through data augmentation and architecture modifications, they continue to struggle with processing long-form text inputs, and the inherent limitations of traditional CLIP text encoders lead to suboptimal downstream generalization. In this paper, we propose FLAME (Frozen Large lAnguage Models Enable data-efficient language-image pre-training) that leverages frozen large language models as text encoders, naturally processing long text inputs and demonstrating impressive multilingual generalization. FLAME comprises two key components: 1) a multifaceted prompt distillation technique for extracting diverse semantic representations from long captions, which better aligns with the multifaceted nature of images, and 2) a facet-decoupled attention mechanism, complemented by an offline embedding strategy, to ensure efficient computation. Extensive empirical evaluations demonstrate FLAME's superior performance. When trained on CC3M, FLAME surpasses the previous state-of-the-art by 4.9% in ImageNet top-1 accuracy. On YFCC15M, FLAME surpasses the WIT-400M-trained CLIP by 44.4% in average image-to-text recall@1 across 36 languages, and by 34.6% in text-to-image recall@1 for long-context retrieval on Urban-1k. Code is available at https://github.com/MIV-XJTU/FLAME.

Dataset distillation has gained popularity as a technique for compressing large datasets into smaller, more efficient representations while retaining essential information for model training. Data features can be broadly divided into two types: instance-specific features, which capture unique, fine-grained details of individual examples, and class-general features, which represent shared, broad patterns across a class. However, previous approaches often struggle to balance these; some focus solely on class-general features, missing finer instance details, while others concentrate on instance-specific features, overlooking the shared characteristics essential for class-level understanding. In this paper, we propose the Non-Critical Region Refinement Dataset Distillation (NRR-DD) method, which preserves the instance-specific and fine-grained regions in synthetic data while enriching non-critical regions with more class-general information. This approach enables our models to leverage all pixel information to capture both types of features, thereby improving overall performance. Furthermore, we introduce Distance-Based Representative (DBR) knowledge transfer, which eliminates the need for soft labels in training by relying solely on the distance between synthetic data predictions and one-hot encoded labels. Experimental results demonstrate that our NRR-DD achieves state-of-the-art performance on both small-scale and large-scale datasets. Additionally, by storing only two distances per instance, our method achieves comparable results across various settings. Code will be available at https://github.com/tmtuan1307/NRR-DD.

Trajectory prediction plays a crucial role in autonomous driving systems, and exploring its vulnerability has garnered widespread attention. However, existing trajectory prediction attack methods often rely on single-point attacks to make efficient perturbations. This limits their applications in real-world scenarios due to the transient nature of single-point attacks, their susceptibility to filtration, and the uncertainty regarding the deployment environment. To address these challenges, this paper proposes a novel LiDAR-induced attack framework to impose multi-frame attacks by optimization-driven adversarial location search, achieving endurance, efficiency, and robustness. This framework strategically places objects near the adversarial vehicle to implement an attack and introduces three key innovations. First, successive state perturbations are generated using a multi-frame single-point attack strategy, effectively misleading trajectory predictions over extended time horizons. Second, we efficiently optimize adversarial objects' locations through three specialized loss functions to achieve desired perturbations. Lastly, we improve robustness by treating the adversarial object as a point without size constraints during the location search phase and reduce dependence on both the specific attack point and the adversarial object's properties. Extensive experiments confirm the superior performance and robustness of our framework.

For the deployment of neural networks in resource-constrained environments, prior works have built lightweight architectures with convolution and attention for capturing local and global dependencies, respectively. Recently, the state space model (SSM) has emerged as an effective operation for global interaction with its favorable linear computational cost in the number of tokens. To harness the efficacy of SSM, we introduce Efficient Vision Mamba (EfficientViM), a novel architecture built on hidden state mixer-based state space duality (HSM-SSD) that efficiently captures global dependencies with further reduced computational cost. With the observation that the runtime of the SSD layer is driven by the linear projections on the input sequences, we redesign the original SSD layer to perform the channel mixing operation within compressed hidden states in the HSM-SSD layer. Additionally, we propose multi-stage hidden state fusion to reinforce the representation power of hidden states and provide the design to alleviate the bottleneck caused by the memory-bound operations. As a result, the EfficientViM family achieves a new state-of-the-art speed-accuracy trade-off on ImageNet-1k, offering up to a 0.7% performance improvement over the second-best model SHViT with faster speed. Further, we observe significant improvements in throughput and accuracy compared to prior works, when scaling images or employing distillation training. Code is available at https://github.com/mlvlab/EfficientViM.

Convolutional neural networks (CNNs) and Vision Transformers (ViTs) have achieved excellent performance in image restoration. While ViTs generally outperform CNNs by effectively capturing long-range dependencies and input-specific characteristics, their computational complexity increases quadratically with image resolution. This limitation hampers their practical application in high-resolution image restoration. In this paper, we propose a simple yet effective visual state space model (EVSSM) for image deblurring, leveraging the benefits of state space models (SSMs) to visual data. In contrast to existing methods that employ several fixed-direction scanning for feature extraction, which significantly increases the computational cost, we develop an efficient visual scan block that applies various geometric transformations before each SSM-based module, capturing useful non-local information and maintaining high efficiency. In addition, to more effectively capture and represent local information, we propose an efficient discriminative frequency domain-based feedforward network (EDFFN) which can effectively estimate useful frequency information for latent clear image restoration. Extensive experimental results show that the proposed EVSSM performs favorably against state-of-the-art methods on benchmark datasets and real-world images.

As a very common type of video, face videos often appear in movies, talk shows, live broadcasts, and other scenes. Real-world online videos are often plagued by degradations such as blurring and quantization noise, due to the high compression ratio caused by high communication costs and limited transmission bandwidth. These degradations have a particularly serious impact on face videos because the human visual system is highly sensitive to facial details.Despite the significant advancement in video face enhancement, current methods still suffer from i) long processing time and ii) inconsistent spatial-temporal visual effects (e.g., flickering). This study proposes a novel and efficient blind video face enhancement method to overcome the above two challenges, restoring high-quality videos from their compressed low-quality versions with an effective de-flickering mechanism. In particular, the proposed method develops upon a 3D-VQGAN backbone associated with spatial-temporal codebooks recording high-quality portrait features and residual-based temporal information. We develop a two-stage learning framework for the model. In Stage I, we learn the model with a regularizer mitigating the codebook collapse problem.In Stage II, we learn two transformers to lookup code from the codebooks and further update the encoder of low-quality videos.Experiments conducted on the VFHQ-Test dataset demonstrate that our method surpasses the current state-of-the-art blind face video restoration and de-flickering methods on both efficiency and effectiveness.

Most current video MLLMs rely on uniform frame sampling and image-level encoders, resulting in inefficient data processing and limited motion awareness. To address these challenges, we introduce EMA, an Efficient Motion-Aware video MLLM that utilizes compressed video structures as inputs. We propose a motion-aware GOP (Group of Pictures) encoder that fuses spatial and motion information within a GOP unit in the compressed video stream, generating compact, informative visual tokens. By integrating fewer but denser RGB frames with more but sparser motion vectors in this native slow-fast input architecture, our approach reduces redundancy and enhances motion representation. Additionally, we introduce MotionBench, a benchmark for evaluating motion understanding across four motion types: linear, curved, rotational, and contact-based. Experimental results show that EMA achieves state-of-the-art performance on both MotionBench and popular video question answering benchmarks, while reducing inference costs. Moreover, EMA demonstrates strong scalability, as evidenced by its competitive performance on long video understanding benchmarks.

Efficient tokenization of videos remains a challenge in training vision models that can process long videos. One promising direction is to develop a tokenizer that can encode long video clips, as it would enable the tokenizer to leverage the temporal coherence of videos better for tokenization. However, training existing tokenizers on long videos often incurs a huge training cost as they are trained to reconstruct all the frames at once. In this paper, we introduce CoordTok, a video tokenizer that learns a mapping from coordinate-based representations to the corresponding patches of input videos, inspired by recent advances in 3D generative models. In particular, CoordTok encodes a video into factorized triplane representations and reconstructs patches that correspond to randomly sampled (x,y,t) coordinates. This allows for training large tokenizer models directly on long videos without requiring excessive training resources. Our experiments show that CoordTok can drastically reduce the number of tokens for encoding long video clips. For instance, CoordTok can encode a 128-frame video with 128x128 resolution into 1280 tokens, while baselines need 6144 or 8192 tokens to achieve similar reconstruction quality. We further show that this efficient video tokenization enables memory-efficient training of a diffusion transformer that can generate 128 frames at once.

Event cameras offer high temporal resolution and dynamic range with minimal motion blur, making them promising for robust object detection. While Spiking Neural Networks (SNNs) on neuromorphic hardware are often considered for energy efficient and low latency event-based data processing, they often fall short of Artificial Neural Networks (ANNs) in accuracy and flexibility. Here, we introduce Attention-based Hybrid SNN-ANN backbones for event-based object detection to leverage the strengths of both SNN and ANN architectures. A novel Attention-based SNN-ANN bridge module proposed to captures sparse spatial and temporal relations from the SNN layer and converts them into dense feature maps for the ANN part of the backbone. Additionally, we present a variant that integrates DWConvLSTMs to the ANN blocks to capture slower dynamics. This multi-timescale network combines fast SNN processing for short timesteps with long-term dense RNN processing, effectively capturing both fast and slow dynamics.Experimental results demonstrate that our proposed method surpasses SNN-based approaches by significant margins, with results comparable to existing ANN and RNN-based methods. Unlike ANN-only networks, the hybrid setup allow us to implement the SNN blocks on digital neuromorphic hardware to investigate the feasibility of our approach.Extensive ablation studies and implementation on neuromorphic hardware confirm the effectiveness of our proposed modules and architectural choices.Our hybrid SNN-ANN architectures pave the way for ANN-like performance at a drastically reduced parameter, latency, and power budget.

Recent 4D dynamic scene editing methods require editing thousands of 2D images used for dynamic scene synthesis and updating the entire scene with additional training loops, resulting in several hours of processing to edit a single dynamic scene. Therefore, these methods are not scalable with respect to the temporal dimension of the dynamic scene (i.e., the number of timesteps). In this work, we propose Instruct-4DGS, an efficient dynamic scene editing method that is more scalable in terms of temporal dimension. To achieve computational efficiency, we leverage a 4D Gaussian representation that models a 4D dynamic scene by combining static 3D Gaussians with a Hexplane-based deformation field, which captures dynamic information. We then perform editing solely on the static 3D Gaussians, which is the minimal but sufficient component required for visual editing. To resolve the misalignment between the edited 3D Gaussians and the deformation field, which may arise from the editing process, we introduce a refinement stage using a score distillation mechanism. Extensive editing results demonstrate that Instruct-4DGS is efficient, reducing editing time by more than half compared to existing methods while achieving high-quality edits that better follow user instructions.

The computational burden of the iterative sampling process remains a major challenge in diffusion-based Low-Light Image Enhancement (LLIE). Current acceleration methods, whether training-based or training-free, often lead to significant performance degradation, highlighting the trade-off between performance and efficiency.In this paper, we identify two primary factors contributing to performance degradation: fitting errors and the inference gap. Our key insight is that fitting errors can be mitigated by linearly extrapolating the incorrect score functions, while the inference gap can be reduced by shifting the Gaussian flow to a reflectance-aware residual space.Based on the above insights, we design Reflectance-Aware Trajectory Refinement (RATR) module, a simple yet effective module to refine the teacher trajectory using the reflectance component of images. Following this, we introduce Reflectance-aware Diffusion with Distilled Trajectory ReDDiT, an efficient and flexible distillation framework tailored for LLIE. Our framework achieves comparable performance to previous diffusion-based methods with redundant steps in just 2 steps while establishing new state-of-the-art (SOTA) results with 8 or 4 steps. Comprehensive experimental evaluations on 10 benchmark datasets validate the effectiveness of our method, consistently outperforming existing SOTA methods.

Stereo depth estimation is a fundamental component in augmented reality (AR), which requires low latency for real-time processing. However, preprocessing such as rectification and non-ML computations such as cost volume require significant amount of latency exceeding that of an ML model itself, which hinders the real-time processing required by AR. Therefore, we develop alternative approaches to the rectification and cost volume that consider ML acceleration (GPU and NPUs) in recent hardware. For pre-processing, we eliminate it by introducing homography matrix prediction network with a rectification positional encoding (RPE), which delivers both low latency and robustness to unrectified images. For cost volume, we replace it with a group-pointwise convolution-based operator and approximation of cosine similarity based on layernorm and dot product. Based on our approaches, we develop MultiHeadDepth (replacing cost volume) and HomoDepth (MultiHeadDepth + removing pre-processing) models. MultiHeadDepth provides 11.8-30.3% improvements in accuracy and 22.9-25.2% reduction in latency compared to a state-of-the-art depth estimation model for AR glasses from industry. HomoDepth, which can directly process unrectified images, reduces the end-to-end latency by 44.5%. We also introduce a multi-task learning method to handle misaligned stereo inputs on HomoDepth, which reduces the AbsRel error by 10.0-24.3%. The overall results demonstrate the efficacy of our approaches, which not only reduce the inference latency but also improve the model performance. Our code is available at https://github.com/UCI-ISA-Lab/MultiHeadDepth-HomoDepth

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.