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.
论文检索
输入标题、作者或关键词，从 2,871 篇学术成果中精准定位
N-shot neural architecture search (NAS) exploits a supernet containing all candidate subnets for a given search space. The subnets are typically trained with a static training strategy (e.g., using the same learning rate (LR) scheduler and optimizer for all subnets). This, however, does not consider that individual subnets have distinct characteristics, leading to two problems: (1) The supernet training is biased towards the low-complexity subnets (unfairness); (2) the momentum update in the supernet is noisy (noisy momentum). We present a dynamic supernet training technique to address these problems by adjusting the training strategy adaptive to the subnets. Specifically, we introduce a complexity-aware LR scheduler (CaLR) that controls the decay ratio of LR adaptive to the complexities of subnets, which alleviates the unfairness problem. We also present a momentum separation technique (MS). It groups the subnets with similar structural characteristics and uses a separate momentum for each group, avoiding the noisy momentum problem. Our approach can be applicable to various N-shot NAS methods with marginal cost, while improving the search performance drastically. We validate the effectiveness of our approach on various search spaces (e.g., NAS-Bench-201, Mobilenet spaces) and datasets (e.g., CIFAR-10/100, ImageNet).
Recent advancements in handwritten text recognition (HTR) have enabled the effective conversion of handwritten text to digital formats. However, achieving robust recognition across diverse writing styles remains challenging. Traditional HTR methods lack writer-specific personalization at test time due to limitations in model architecture and training strategies. Existing attempts to bridge this gap, through gradient-based meta-learning, still require labeled examples and suffer from parameter-inefficient fine-tuning, leading to substantial computational and memory overhead. To overcome these challenges, we propose an efficient framework that formulates personalization as prompt tuning, incorporating an auxiliary image reconstruction task with a self-supervised loss to guide prompt adaptation with unlabeled test-time examples. To ensure self-supervised loss effectively minimizes text recognition error, we leverage meta-learning to learn the optimal initialization of the prompts. As a result, our method allows the model to efficiently capture unique writing styles by updating less than 1% of its parameters and eliminating the need for time-intensive annotation processes. We validate our approach on the RIMES and IAM Handwriting Database benchmarks, where it consistently outperforms previous state-of-the-art methods while using 20x fewer parameters. We believe this represents a significant advancement in personalized handwritten text recognition, paving the way for more reliable and practical deployment in resource-constrained scenarios.
Task Incremental Learning (TIL) is a specialized form of Continual Learning (CL) in which a model incrementally learns from non-stationary data streams. Existing TIL methodologies operate under the closed-world assumption, presuming that incoming data remains in-distribution (ID). However, in an open-world setting, incoming samples may originate from out-of-distribution (OOD) sources, with their task identities inherently unknown. Continually detecting OOD samples presents several challenges for current OOD detection methods: reliance on model outputs leads to excessive dependence on model performance, selecting suitable thresholds is difficult, hindering real-world deployment, and binary ID/OOD classification fails to provide task-level identification. To address these issues, we propose a novel continual OOD detection method called the Hierarchical Two-sample Tests (H2ST). H2ST eliminates the need for threshold selection through hypothesis testing and utilizes feature maps to better exploit model capabilities without excessive dependence on model performance. The proposed hierarchical architecture enables task-level detection with superior performance and lower overhead compared to non-hierarchical classifier two-sample tests. Extensive experiments and analysis validate the effectiveness of H2ST in open-world TIL scenarios and its superiority to the existing methods. Code is available at https://github.com/YuhangLiuu/H2ST.
We present Pippo, a generative model capable of producing 1K resolution dense turnaround videos of a person from a single casually clicked photo. Pippo is a multi-view diffusion transformer and does not require any additional inputs - e.g., a fitted parametric model or camera parameters of the input image. We pre-train Pippo on 3B human images without captions, and conduct multi-view mid-training and post-training on studio captured humans. During mid-training, to quickly absorb the studio dataset, we denoise several (up to 48) views at low-resolution, and encode target cameras coarsely using a shallow MLP. During post-training, we denoise fewer views at high-resolution and use pixel-aligned controls (e.g., Spatial anchor and Plucker rays) to enable 3D consistent generations. At inference, we propose an attention biasing technique that allows Pippo to simultaneously generate greater than 5 times as many views as seen during training. Finally, we also introduce an improved metric to evaluate 3D consistency of multi-view generations, and show that Pippo outperforms existing works on multi-view human generation from a single image.
Lessons and Insights from a Unifying Study of Parameter-Efficient Fine-Tuning (PEFT) in Visual Recognition
PDF ↗Parameter-efficient fine-tuning (PEFT) has attracted significant attention due to the growth of pre-trained model sizes and the need to fine-tune (FT) them for superior downstream performance. Despite a surge in new PEFT methods, a systematic study to understand their performance and suitable application scenarios is lacking, leaving questions like "when to apply PEFT" and "which method to use" largely unanswered, especially in visual recognition. In this paper, we conduct a unifying empirical study of representative PEFT methods with Vision Transformers. We systematically tune their hyperparameters to fairly compare their accuracy on downstream tasks. Our study offers a practical user guide and unveils several new insights. First, if tuned carefully, different PEFT methods achieve similar accuracy in the low-shot benchmark VTAB-1K. This includes simple approaches like FT the bias terms that were reported inferior. Second, despite similar accuracy, we find that PEFT methods make different mistakes and high-confidence predictions, likely due to their different inductive biases. Such an inconsistency (or complementarity) opens up the opportunity for ensemble methods, and we make preliminary attempts at this. Third, going beyond the commonly used low-shot tasks, we find that PEFT is also useful in many-shot regimes, achieving comparable or better accuracy than full FT while using significantly fewer parameters. Lastly, we investigate PEFT's ability to preserve a pre-trained model's robustness to distribution shifts (e.g., CLIP). Perhaps not surprisingly, PEFT approaches outperform full FT alone. However, with weight-space ensembles, full FT can better balance target distribution and distribution shift performance, suggesting a future research direction for robust PEFT. The code is available at https://github.com/OSU-MLB/ViT_PEFT_Vision.
Recent advances in large-scale text-to-image (T2I) diffusion models have enabled a variety of downstream applications. As T2I models require extensive resources for training, they constitute highly valued intellectual property (IP) for their legitimate owners, yet making them incentive targets for unauthorized fine-tuning by adversaries seeking to leverage these models for customized, usually profitable applications. Existing IP protection methods for diffusion models generally involve embedding watermark patterns and then verifying ownership through generated outputs examination, or inspecting the model's feature space. However, these techniques are inherently ineffective in practical scenarios when the watermarked model undergoes fine-tuning, and the feature space is inaccessible during verification (i.e., black-box setting). The model is prone to forgetting the previously learned watermark knowledge when it adapts to a new task. To address this challenge, we propose SleeperMark, a novel framework designed to embed resilient watermarks into T2I diffusion models. SleeperMark explicitly guides the model to disentangle the watermark information from the semantic concepts it learns, allowing the model to retain the embedded watermark while continuing to be adapted to new downstream tasks. Our extensive experiments demonstrate the effectiveness of SleeperMark across various types of diffusion models, including latent diffusion models (e.g., Stable Diffusion) and pixel diffusion models (e.g., DeepFloyd-IF), showing robustness against downstream fine-tuning and various attacks at both the image and model levels, with minimal impact on the model's generative capability.
We present LoTUS, a novel Machine Unlearning (MU) method that eliminates the influence of training samples from pre-trained models, avoiding retraining from scratch. LoTUS smooths the prediction probabilities of the model up to an information-theoretic bound, mitigating its over-confidence stemming from data memorization. We evaluate LoTUS on Transformer and ResNet18 models against eight baselines across five public datasets. Beyond established MU benchmarks, we evaluate unlearning on ImageNet1k, a large-scale dataset, where retraining is impractical, simulating real-world conditions. Moreover, we introduce the novel Retrain-Free Jensen-Shannon Divergence (RF-JSD) metric to enable evaluation under real-world conditions. The experimental results show that LoTUS outperforms state-of-the-art methods in terms of both efficiency and effectiveness. Code: https://github.com/cspartalis/LoTUS
As AI models are increasingly integrated into applications involving human interaction, understanding the alignment between human perception and machine vision has become essential. One example is the estimation of visual motion (optical flow) in dynamic applications such as driving assistance. While there are numerous optical flow datasets and benchmarks with ground truth information, human-perceived flow in natural scenes remains underexplored. We introduce HuPerFlow--a benchmark for human-perceived flow, measured at 2,400 locations across ten optical flow datasets, with 38,400 response vectors collected through online psychophysical experiments. Our data demonstrate that human-perceived flow aligns with ground truth in spatiotemporally smooth locations while also showing systematic errors influenced by various environmental properties. Additionally, we evaluated several optical flow algorithms against human-perceived flow, uncovering both similarities and unique aspects of human perception in complex natural scenes. HuPerFlow is the first large-scale human-perceived flow benchmark for alignment between computer vision models and human perception, as well as for scientific exploration of human motion perception in natural scenes. The HuPerFlow benchmark will be available online upon acceptance.
Recently, zero-shot methods like InstantID have revolutionized identity-preserving generation. Unlike multi-image finetuning approaches such as DreamBooth, these zero-shot methods leverage powerful facial encoders to extract identity information from a single portrait photo, enabling efficient identity-preserving generation through a single inference pass. However, this convenience introduces new threats to the facial identity protection. This paper aims to safeguard portrait photos from unauthorized encoder-based customization. We introduce IDProtector, an adversarial noise encoder that applies imperceptible adversarial noise to portrait photos in a single forward pass. Our approach offers universal protection for portraits against multiple state-of-the-art encoder-based methods, including InstantID, IP-Adapter, and PhotoMaker, while ensuring robustness to common image transformations such as JPEG compression, resizing, and affine transformations. Experiments across diverse portrait datasets and generative models reveal that IDProtector generalizes effectively to unseen data and even closed-source proprietary models.
Seeing is Not Believing: Adversarial Natural Object Optimization for Hard-Label 3D Scene Attacks
PDF ↗Deep learning models for 3D data have shown to be vulnerable to adversarial attacks, which have received increasing attention in various safety-critical applications such as autonomous driving and robotic navigation. Existing 3D attackers mainly put effort into attacking the simple 3D classification model by perturbing point cloud objects in the white/black-box setting. However, real-world 3D applications focus on tackling more complicated scene-based data while sharing no information about the model parameters and logits with users. Therefore, directly applying previous naive 3D attack methods to these applications does not work. To this end, this paper attempts to address the challenging hard-label 3D scene attack with access only to the input/output of the 3D models. To make the attack effective and stealthy, we propose to generate universal adversarial objects, which will mislead scene-aware 3D models to predict attacker-chosen labels whenever these objects are placed on any scene input. Specifically, we inject an imperceptible object trigger with further perturbations into all scenes and learn to mislead their reasoning by only querying the 3D model. We start by initializing the trigger pattern with a realistic object and searching for an appropriate location to place it naturally in the scene data. Then, we design a novel weighted gradient estimation strategy to perturb the object trigger with additive slight noise to make them adversarial in an iterative optimization procedure. Extensive experiments demonstrate that our attack can achieve superior performance on seven 3D models and three scene-based datasets, with satisfactory adversarial imperceptibility and strong resistance to defense methods.
A fundamental aspect of compositional reasoning in a video is associating people and their actions across time. Recent years have seen great progress in general-purpose vision/video models and a move towards long-video understanding. While exciting, we take a step back and ask: are today's models good at compositional reasoning on short videos? To this end, we introduce VELOCITI, a benchmark to study Video-LLMs by disentangling and assessing the comprehension of agents, actions, and their associations across multiple events. We adopt the Video-Language Entailment setup and propose StrictVLE that requires correct classification (rather than ranking) of the positive and negative caption. We evaluate several models and observe that even the best, LLaVA-OneVision (44.5%) and Gemini-1.5-Pro (49.3%), are far from human accuracy at 93.0%. Results show that action understanding lags behind agents, and negative captions created using entities appearing in the video perform worse than those obtained from pure text manipulation. We also present challenges with ClassicVLE and multiple-choice (MC) evaluation, strengthening our preference for StrictVLE. Finally, we validate that our benchmark requires visual inputs of multiple frames making it ideal to study video-language compositional reasoning.
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.
nnWNet: Rethinking the Use of Transformers in Biomedical Image Segmentation and Calling for a Unified Evaluation Benchmark
PDF ↗Semantic segmentation is a crucial prerequisite in clinical applications and computer-aided diagnosis. With the development of deep neural networks, biomedical image segmentation has achieved remarkable success. Encoder-decoder architectures that integrate convolutions and transformers are gaining attention for their potential to capture both global and local features. However, current designs face the contradiction that these two features cannot be continuously transmitted. In addition, some models lack a unified and standardized evaluation benchmark, leading to significant discrepancies in the experimental setup. In this study, we review and summarize these architectures and analyze their contradictions in design. We modify UNet and propose WNet to combine transformers and convolutions, addressing the transmission issue effectively. WNet captures long-range dependencies and local details simultaneously while ensuring their continuous transmission and multi-scale fusion. We integrate WNet into the nnUNet framework for unified benchmarking. Our model achieves state-of-the-art performance in biomedical image segmentation. Extensive experiments demonstrate their effectiveness on four 2D datasets (DRIVE, ISIC-2017, Kvasir-SEG, and CREMI) and four 3D datasets (Parse2022, AMOS22, BTCV, and ImageCAS). The code is available at https://github.com/Yanfeng-Zhou/nnWNet.
Autonomous driving evaluation requires simulation environments that closely replicate actual road conditions, including real-world sensory data and responsive feedback loops. However, many existing simulations need to predict waypoints along fixed routes on public datasets or synthetic photorealistic data, i.e., open-loop simulation usually lacks the ability to assess dynamic decision-making. While the recent efforts of closed-loop simulation offer feedback-driven environments, they cannot process visual sensor inputs or produce outputs that differ from real-world data. To address these challenges, we propose DrivingSphere, a realistic and closed-loop simulation framework. Its core idea is to build 4D world representation and generate real-life and controllable driving scenarios. In specific, our framework includes a Dynamic Environment Composition module that constructs a detailed 4D driving world with a format of occupancy equipping with static backgrounds and dynamic objects, and a Visual Scene Synthesis module that transforms this data into high-fidelity, multi-view video outputs, ensuring spatial and temporal consistency. By providing a dynamic and realistic simulation environment, DrivingSphere enables comprehensive testing and validation of autonomous driving algorithms, ultimately advancing the development of more reliable autonomous cars.The benchmark will be publicly released.
SDGOCC: Semantic and Depth-Guided Bird's-Eye View Transformation for 3D Multimodal Occupancy Prediction
PDF ↗Multimodal 3D occupancy prediction has garnered significant attention for its potential in autonomous driving. However, most existing approaches are single-modality: camera-based methods lack depth information, while LiDAR-based methods struggle with occlusions. Current lightweight methods primarily rely on the Lift-Splat-Shoot (LSS) pipeline, which suffers from inaccurate depth estimation and fails to fully exploit the geometric and semantic information of 3D LiDAR points. Therefore, we propose a novel multimodal occupancy prediction network called SDG-OCC, which incorporates a joint semantic and depth-guided view transformation coupled with a fusion-to-occupancy-driven active distillation. The enhanced view transformation constructs accurate depth distributions by integrating pixel semantics and co-point depth through diffusion and bilinear discretization. The fusion-to-occupancy-driven active distillation extracts rich semantic information from multimodal data and selectively transfers knowledge to image features based on LiDAR-identified regions. Finally, for optimal performance, we introduce SDG-Fusion, which uses fusion alone, and SDG-KL, which integrates both fusion and distillation for faster inference. Our method achieves state-of-the-art (SOTA) performance with real-time processing on the Occ3D-nuScenes dataset and shows comparable performance on the more challenging SurroundOcc-nuScenes dataset, demonstrating its effectiveness and robustness. The code will be released at https://github.com/DzpLab/SDGOCC.
Drones have become essential tools for reconstructing wild scenes due to their outstanding maneuverability. Recent advances in radiance field methods have achieved remarkable rendering quality, providing a new avenue for 3D reconstruction from drone imagery. However, dynamic distractors in wild environments challenge the static scene assumption in radiance fields, while limited view constraints hinder the accurate capture of underlying scene geometry. To address these challenges, we introduce DroneSplat, a novel framework designed for robust 3D reconstruction from in-the-wild drone imagery. Our method adaptively adjusts masking thresholds by integrating local-global segmentation heuristics with statistical approaches, enabling precise identification and elimination of dynamic distractors in static scenes. We enhance 3D Gaussian Splatting with multi-view stereo predictions and a voxel-guided optimization strategy, supporting high-quality rendering under limited view constraints. For comprehensive evaluation, we provide a drone-captured 3D reconstruction dataset encompassing both dynamic and static scenes. Extensive experiments demonstrate that DroneSplat outperforms both 3DGS and NeRF baselines in handling in-the-wild drone imagery.
Fast and reliable omnidirectional 3D sensing is essential to many applications such as autonomous driving, robotics and drone navigation. While many well-recognized methods have been developed to produce high-quality omnidirectional 3D information, they are too slow for real-time computation, limiting their feasibility in practical applications. Motivated by these shortcomings, we propose an efficient omnidirectional depth sensing framework, called OmniStereo, which generates high-quality 3D information in real-time. Unlike prior works, OmniStereo employs Cassini projection to simplify the photometric matching and introduces a lightweight stereo matching network to minimize computational overhead. Additionally, OmniStereo proposes a novel fusion method to handle depth discontinuities and invalid pixels complemented by a refinement module to reduce mapping-introduced errors and recover fine details. As a result, OmniStereo achieves state-of-the-art (SOTA) accuracy, surpassing the second-best method over 32% in MAE, while maintaining real-time efficiency. It operates more than 16.5xfaster than the second-best method in accuracy on TITAN RTX, achieving 12.3 FPS on embedded device Jetson AGX Orin, underscoring its suitability for real-world deployment. The code is available at https://github.com/DengJiaxi1/OmniStereo.
Deep neural networks (DNNs) exhibit vulnerability to adversarial examples that can transfer across different DNN models. A particularly challenging problem is developing transferable targeted attacks that can mislead DNN models into predicting specific target classes. While various methods have been proposed to enhance attack transferability, they often incur substantial computational costs while yielding limited improvements. Recent clean feature mixup methods use random clean features to perturb the feature space but lack optimization for disrupting adversarial examples, overlooking the advantages of attack-specific perturbations. In this paper, we propose Feature Tuning Mixup (FTM), a novel method that enhances targeted attack transferability by combining both random and optimized noises in the feature space. FTM introduces learnable feature perturbations and employs an efficient stochastic update strategy for optimization. These learnable perturbations facilitate the generation of more robust adversarial examples with improved transferability. We further demonstrate that attack performance can be enhanced through an ensemble of multiple FTM-perturbed surrogate models. Extensive experiments on the ImageNet-compatible dataset across various DNN models demonstrate that our method achieves significant improvements over state-of-the-art methods while maintaining low computational cost.
Text-to-image (T2I) diffusion models have revolutionized visual content creation, but extending these capabilities to text-to-video (T2V) generation remains a challenge, particularly in preserving temporal consistency. Existing methods that aim to improve consistency often cause trade-offs such as reduced imaging quality and impractical computational time. To address these issues we introduce VideoGuide, a novel framework that enhances the temporal consistency of pretrained T2V models without the need for additional training or fine-tuning. Instead, VideoGuide leverages any pretrained video diffusion model (VDM) or itself as a guide during the early stages of inference, improving temporal quality by interpolating the guiding model's denoised samples into the sampling model's denoising process. The proposed method brings about significant improvement in temporal consistency and image fidelity, providing a cost-effective and practical solution that synergizes the strengths of various video diffusion models. Furthermore, we demonstrate prior distillation, revealing that base models can achieve enhanced text coherence by utilizing the superior data prior of the guiding model through the proposed method. Project Page: https://dohunlee1.github.io/videoguide.github.io/