Temporal Action Detection (TAD) is fundamental yet challenging for real-world video applications. Leveraging the unique benefits of transformers, various DETR-based approaches have been adopted in TAD. However, it has recently been identified that the attention collapse in self-attention causes the performance degradation of DETR for TAD. Building upon previous research, this paper newly addresses the attention collapse problem in cross-attention within DETR-based TAD methods. Moreover, our findings reveal that cross-attention exhibits patterns distinct from predictions, indicating a short-cut phenomenon. To resolve this, we propose a new framework, Prediction-Feedback DETR (Pred-DETR), which utilizes predictions to restore the collapse and align the cross- and self-attention with predictions. Specifically, we devise novel prediction-feedback objectives using guidance from the relations of the predictions. As a result, Pred-DETR significantly alleviates the collapse and achieves state-of-the-art performance among DETR-based methods on various challenging benchmarks including THUMOS14, ActivityNet-v1.3, HACS, and FineAction.
论文检索
输入标题、作者或关键词，从 8,216 篇学术成果中精准定位
Foundational vision-language models like CLIP are emerging as a promising paradigm in vision due to their excellent generalization. However, adapting these models for downstream tasks while maintaining their generalization remains challenging. In literature, one branch of methods adapts CLIP by learning prompts using images. While effective, these methods often rely on image-label data, which is not always practical, and struggle to generalize to new datasets due to overfitting on few-shot source data. Another approach explores training-free methods by generating class captions from large language models (LLMs) and performing prompt ensembling, but these methods often produce static, class-specific prompts that cannot be transferred to new classes and incur additional costs by generating LLM descriptions for each class separately. In this work, we aim to combine the strengths of both approaches by learning prompts using only text data derived from LLMs. As supervised training of prompts in the image-free setup is non-trivial, we develop a language-only efficient training approach that enables prompts to distill rich contextual knowledge from LLM data. Furthermore, by mapping the LLM contextual text data within the learned prompts, our approach enables zero-shot transfer of prompts to new classes and datasets, potentially reducing the LLM prompt engineering cost. To the best of our knowledge, this is the first work that learns generalized and transferable prompts for image tasks using only text data. We perform evaluations on 4 benchmarks, where ProText improves over ensembling methods while being competitive with those using labeled images.
Neural Radiance Fields (NeRF) have achieved huge success in effectively capturing and representing 3D objects and scenes. However, to establish an ubiquitous presence in everyday media formats, such as images and videos, we need to fulfill three key objectives: 1. fast encoding and decoding time, 2. compact model sizes, and 3. high-quality renderings. Despite recent advancements, a comprehensive algorithm that adequately addresses all objectives has yet to be fully realized. In this work, we present CodecNeRF, a neural codec for NeRF representations, consisting of an encoder and decoder architecture that can generate a NeRF representation in a single forward pass. Furthermore, inspired by the recent parameter-efficient finetuning approaches, we propose a finetuning method to efficiently adapt the generated NeRF representations to a new test instance, leading to high-quality image renderings and compact code sizes. The proposed CodecNeRF, a newly suggested encoding-decoding-finetuning pipeline for NeRF, achieved unprecedented compression performance of more than 100x and a remarkable reduction in encoding time while maintaining (or improving) the image quality on widely used 3D object datasets.
Pedestrian Attribute Recognition: A New Benchmark Dataset and a Large Language Model Augmented Framework
PDF ↗Pedestrian Attribute Recognition (PAR) is one of the indispensable tasks in human-centered research. However, existing datasets neglect different domains (e.g., environments, times, populations, and data sources), only conducting simple random splits, and the performance of these datasets has already approached saturation. In the past five years, no large-scale dataset has been opened to the public. To address this issue, this paper proposes a new large-scale, cross-domain pedestrian attribute recognition dataset to fill the data gap, termed MSP60K. It consists of 60,122 images and 57 attribute annotations across eight scenarios. Synthetic degradation is also conducted to further narrow the gap between the dataset and real-world challenging scenarios. To establish a more rigorous benchmark, we evaluate 17 representative PAR models under both random and cross-domain split protocols on our dataset. Additionally, we propose an innovative Large Language Model (LLM) augmented PAR framework, named LLM-PAR. This framework processes pedestrian images through a Vision Transformer (ViT) backbone to extract features and introduces a multi-embedding query Transformer to learn partial-aware features for attribute classification. Significantly, we enhance this framework with LLM for ensemble learning and visual feature augmentation. Comprehensive experiments across multiple PAR benchmark datasets have thoroughly validated the efficacy of our proposed framework.
The rapid development of large-scale deep learning models questions the affordability of hardware platforms, which necessitates the pruning to reduce their computational and memory footprints. Sparse neural networks as the product, have demonstrated numerous favorable benefits like low complexity, undamaged generalization, etc. Most of the prominent pruning strategies are invented from a model-centric perspective, focusing on searching and preserving crucial weights by analyzing network topologies. However, the role of data and its interplay with model-centric pruning has remained relatively unexplored. In this research, we introduce a novel data-model co-design perspective: to promote superior weight sparsity by learning important model topology and adequate input data in a synergetic manner. Specifically, customized Visual Prompts are mounted to upgrade neural Network sparsification in our proposed VPNs framework. As a pioneering effort, this paper conducts systematic investigations about the impact of different visual prompts on model pruning and suggests an effective joint optimization approach. Extensive experiments with 3 network architectures and 8 datasets evidence the substantial performance improvements from VPNs over existing start-of-the-art pruning algorithms. Furthermore, we find that subnetworks discovered by VPNs from pre-trained models enjoy better transferability across diverse downstream scenarios. These insights shed light on new promising possibilities of data-model co-designs for vision model sparsification.
Human pose estimation in videos remains a challenge, largely due to the reliance on extensive manual annotation of large datasets, which is expensive and labor-intensive. Furthermore, existing approaches often struggle to capture long-range temporal dependencies and overlook the complementary relationship between temporal pose heatmaps and visual features. To address these limitations, we introduce STDPose, a novel framework that enhances human pose estimation by learning spatiotemporal dynamics in sparsely-labeled videos. STDPose incorporates two key innovations: 1) A novel Dynamic-Aware Mask to capture long-range motion context, allowing for a nuanced understanding of pose changes. 2) A system for encoding and aggregating spatiotemporal representations and motion dynamics to effectively model spatiotemporal relationships, improving the accuracy and robustness of pose estimation. STDPose establishes a new performance benchmark for both video pose propagation (i.e., propagating pose annotations from labeled frames to unlabeled frames) and pose estimation tasks, across three large-scale evaluation datasets. Additionally, utilizing pseudo-labels generated by pose propagation, STDPose achieves competitive performance with only 26.7% labeled data.
Instance segmentation algorithms in remote sensing are typically based on conventional methods, limiting their application to seen scenarios and closed-set predictions. In this work, we propose a novel task called zero-shot remote sensing instance segmentation, aimed at identifying aerial objects that are absent from training data. Challenges arise when classifying aerial categories with high inter-class similarity and intra-class variance. Besides, the domain gap between vision-language models’ pretraining datasets and remote sensing datasets hinders the zero-shot capabilities of the pretrained model when it is directly applied to remote sensing images. To address these challenges, we propose a Zero-Shot Remote Sensing Instance Segmentation framework, dubbed ZoRI. Our approach features a discrimination-enhanced classifier that uses refined textual embeddings to increase the awareness of class disparities. Instead of direct fine-tuning, we propose a knowledge-maintained adaptation strategy that decouples semantic-related information to preserve vision-language alignment while adjusting features to capture remote sensing domain-specific visual cues. Additionally, we introduce a prior-injected prediction with cache bank of aerial visual prototypes to supplement the semantic richness of text embeddings and seamlessly integrate aerial representations, adapting to the remote sensing domain. We establish new experimental protocols and benchmarks, and extensive experiments demonstrate that ZoRI achieves the state-of-art performance on the zero-shot remote sensing instance segmentation task.
Dynamic object modeling is a critical challenge in 3D scene reconstruction. Previous methods typically maintain a canonical space to represent the object model, and a deformation field to express the object motion. However, this approach fails when the object undergoes large motions. The position variation caused by significant motion not only complicates the establishment of a canonical space, but also misleads the interpretation of the deformation field. To overcome the above weaknesses, we propose Motion Decoupled Dynamic 3D Gaussian Splatting (M5D-GS), the first 3D-GS model that separates motion and deformation modeling for dynamic object representation with large motion from a monocular camera. M5D-GS increases the practicality of 3D-GS, as it is common for objects to move, rotate, and deform simultaneously. Current datasets only contain object deformations with slight motions. We introduce a pipeline to reuse current benchmarks by adding large motions into the scene. We also introduce a new benchmark featuring several new scenes with complex motions, scenes augmented from previous datasets, and some real world recorded testcases, to fully demonstrate our improvements. Our M5D-GS significantly increases the accuracy under large motion scenarios while maintaining high rendering speed, which makes it suitable for dynamic object representation tasks including 4D novel view synthesis and real-time rendering.
Training-and-Prompt-Free General Painterly Harmonization via Zero-Shot Disentenglement on Style and Content References
PDF ↗Painterly image harmonization aims at seamlessly blending disparate visual elements within a single image. However, previous approaches often struggle due to limitations in training data or reliance on additional prompts, leading to inharmonious and content-disrupted output. To surmount these hurdles, we design a Training-and-prompt-Free General Painterly Harmonization method (TF-GPH). TF-GPH incorporates a novel “Similarity Disentangle Mask”, which disentangles the foreground content and background image by redirecting their attention to corresponding reference images, enhancing the attention mechanism for multi-image inputs. Additionally, we propose a “Similarity Reweighting” mechanism to balance harmonization between stylization and content preservation. This mechanism minimizes content disruption by prioritizing the content-similar features within the given background style reference. Finally, we address the deficiencies in existing benchmarks by proposing novel range-based evaluation metrics and a new benchmark to better reflect real-world applications. Extensive experiments demonstrate the efficacy of our method across benchmarks.
Recent advancements in online Video Instance Segmentation (VIS) methods show notable performance improvements across benchmarks. However, the leading methods in the tracking-by-detection paradigm often result in temporally inconsistent predictions at both instance-level and pixel-level that lead to visually unsatisfactory outcomes. To address these challenges, we propose RoCoVIS, a simple yet effective approach that integrates segmentation and tracking to provide consistent online VIS. Our approach is an end-to-end sequential learning where object queries are propagated through mask predictions, improving the accuracy of temporal instance mapping at the pixel level. Additionally, we propose a new label assignment criterion in harmony with our approach. We also examine the limitations and challenges presented by the current standard evaluation protocol (AP) and suggest adopting additional metrics, Tube-Boundary AP and AP_Pool. RoCoVIS demonstrates superior performance on challenging VIS benchmarks with a Swin-L backbone and shows competitive results when employing a ResNet-50 backbone. By employing Tube-Boundary AP and AP_Pool as metrics to measure mask accuracy and consistency, RoCoVIS outperforms its counterpart, GenVIS, on the HQ-YTVIS and VIPSeg.
The recent Segment Anything Model (SAM) has emerged as a new paradigmatic vision foundation model, showcasing potent zero-shot generalization and flexible prompting. Despite SAM finding applications and adaptations in various domains, its primary limitation lies in the inability to grasp object semantics. In this paper, we present Sambor to seamlessly integrate SAM with the open-vocabulary object detector in an end-to-end framework. While retaining all the remarkable capabilities inherent to SAM, we boost it to detect arbitrary objects from human inputs like category names or reference expressions. Building upon the SAM image encoder, we introduce a novel SideFormer module designed to acquire SAM features adept at perceiving objects and inject comprehensive semantic information for recognition. In addition, we devise an Open-set RPN that leverages SAM proposals to assist in finding potential objects. Consequently, Sambor enables the open-vocabulary detector to equally focus on generalizing both localization and classification sub-tasks. Our approach demonstrates superior zero-shot performance across benchmarks, including COCO and LVIS, proving highly competitive against previous state-of-the-art methods. We aspire for this work to serve as a meaningful endeavor in endowing SAM to recognize diverse object categories and advancing open-vocabulary learning with the support of vision foundation models.
Neural Representations for Videos (NeRV) has emerged as a promising implicit neural representation (INR) approach for video analysis, which represents videos as neural networks with frame indexes as inputs. However, NeRV-based methods are time-consuming when adapting to a large number of diverse videos, as each video requires a separate NeRV model to be trained from scratch. In addition, NeRV-based methods spatially require generating a high-dimension signal (i.e., an entire image) from the input of a low-dimension timestamp, and a video typically consists of tens of frames temporally that have a minor change between adjacent frames. To improve the efficiency of video representation, we propose Meta Neural Representations for Videos, named MetaNeRV, a novel framework for fast NeRV representation for unseen videos. MetaNeRV leverages a meta-learning framework to learn an optimal parameter initialization, which serves as a good starting point for adapting to new videos. To address the unique spatial and temporal characteristics of video modality, we further introduce spatial-temporal guidance to improve the representation capabilities of MetaNeRV. Specifically, the spatial guidance with a multi-resolution loss aims to capture the information from different resolution stages, and the temporal guidance with an effective progressive learning strategy could gradually refine the number of fitted frames during the meta-learning process. Extensive experiments conducted on multiple datasets demonstrate the superiority of MetaNeRV for video representations and video compression.
The generation of a virtual digital avatar is a crucial research topic in the field of computer vision. Many existing works utilize Neural Radiance Fields (NeRF) to address this issue and have achieved impressive results. However, previous works assume the images of the training person are available and fixed while the appearances and poses of a subject could constantly change and increase in real-world scenarios. How to update the human avatar but also maintain the ability to render the old appearance of the person is a practical challenge. One trivial solution is to combine the existing virtual avatar models based on NeRF with continual learning methods. However, there are some critical issues in this approach: learning new appearances and poses can cause the model to forget past information, which in turn leads to a degradation in the rendering quality of past appearances, especially color bleeding issues, and incorrect human body poses. In this work, we propose a maintainable avatar (MaintaAvatar) based on neural radiance fields by continual learning, which resolves the issues by utilizing a Global-Local Joint Storage Module and a Pose Distillation Module. Overall, our model requires only limited data collection to quickly fine-tune the model while avoiding catastrophic forgetting, thus achieving a maintainable virtual avatar. The experimental results validate the effectiveness of our MaintaAvatar model.
In recent years, deep learning has revenue in automated medical landmark detection. Nonetheless, prevailing research in this field predominantly addresses single-center scenarios or domain adaptation settings. In practical environments, the acquisition of multi-center data faces privacy concerns, coupled with the time-intensive and costly nature of data collection and annotation. These challenges substantially impede the broader application of deep learning-based medical landmark detection. To mitigate these issues, we propose a novel domain-generalized medical landmark detection framework that relies solely on single-center data for training. Considering the availability of numerous public medical segmentation datasets, we design a simple yet effective method that utilizes single-center segmentation to enhance the domain generalization capabilities of the landmark detection task. Specifically, we introduce a novel boundary-aware pre-training approach to focus the model on regions pertinent to landmarks. To further enhance the robustness and generalization capabilities during pre-training, we have derived a mixing loss term and proved its effectiveness in theory and practice. Extensive experiments conducted on our new domain generalization benchmark for medical landmark detection demonstrate the superiority of our approach.
EventMamba: Enhancing Spatio-Temporal Locality with State Space Models for Event-Based Video Reconstruction
PDF ↗Leveraging its robust linear global modeling capability, Mamba has notably excelled in computer vision. Despite its success, existing Mamba-based vision models have overlooked the nuances of event-driven tasks, especially in video reconstruction. Event-based video reconstruction (EBVR) demands spatial translation invariance and close attention to local event relationships in the spatio-temporal domain. Unfortunately, conventional Mamba algorithms apply static window partitions and standard reshape scanning methods, leading to significant losses in local connectivity. To overcome these limitations, we introduce EventMamba—a specialized model designed for EBVR task. EventMamba innovates by incorporating random window offset (RWO) in the spatial domain, moving away from the restrictive fixed partitioning. Additionally, it features a new consistent traversal serialization approach in the spatio-temporal domain, which maintains the proximity of adjacent events both spatially and temporally. These enhancements enable EventMamba to retain Mamba’s robust modeling capabilities while significantly preserving the spatio-temporal locality of event data. Comprehensive testing on multiple datasets shows that EventMamba markedly enhances video reconstruction, drastically improving computation speed while delivering superior visual quality compared to Transformer-based methods.
TC-LLaVA: Rethinking the Transfer of LLava from Image to Video Understanding with Temporal Considerations
PDF ↗Multimodal Large Language Models (MLLMs) have significantly improved performance across various image-language applications. Recently, there has been a growing interest in adapting image pre-trained MLLMs for video-related tasks. However, most efforts concentrate on enhancing the vision encoder and projector components, while the core part, Large Language Models (LLMs), remains comparatively under-explored. In this paper, we propose two strategies to enhance the model's capability in video understanding tasks by improving inter-layer attention computation in LLMs. Specifically, the first approach focuses on the enhancement of Rotary Position Embedding (RoPE) with Temporal-Aware Dual RoPE, which introduces temporal position information to strengthen the MLLM's temporal modeling capabilities while preserving the relative position relationships of both visual and text tokens. The second approach involves enhancing the Attention Mask with the Frame-wise Block Causal Attention Mask, a simple yet effective method that broadens visual token interactions within and across video frames while maintaining the causal inference mechanism. Based on these proposed methods, we adapt LLaVA for video understanding tasks, naming it Temporal-Considered LLaVA (TC-LLaVA). Our TC-LLaVA achieves new state-of-the-art performance across various video understanding benchmarks with only supervised fine-tuning (SFT) on video-related datasets.
Neural video compression has recently demonstrated significant potential to compete with conventional video codecs in terms of rate-quality performance. These learned video codecs are however associated with various issues related to decoding complexity (for autoencoder-based methods) and/or system delays (for implicit neural representation (INR) based models), which currently prevent them from being deployed in practical applications. In this paper, targeting a practical neural video codec, we propose a novel INR-based coding framework, PNVC, which innovatively combines autoencoder-based and overfitted solutions. Our approach benefits from several design innovations, including a new structural reparameterization-based architecture, hierarchical quality control, modulation-based entropy modeling, and scale-aware positional embedding. Supporting both low delay (LD) and random access (RA) configurations, PNVC outperforms existing INR-based codecs, achieving nearly 35%+ BD-rate savings against HEVC HM 18.0 (LD) - almost 10% more compared to one of the state-of-the-art INR-based codecs, HiNeRV and 5% more over VTM 20.0 (LD), while maintaining 20+ FPS decoding speeds for 1080p content. This represents an important step forward for INR-based video coding, moving it towards practical deployment.
It is widely known that state-of-the-art machine learning models, including vision and language models, can be seriously compromised by adversarial perturbations. It is therefore increasingly relevant to develop capabilities to certify their performance in the presence of the most effective adversarial attacks. Our paper offers a new approach to certify the performance of machine learning models in the presence of adversarial attacks with population level risk guarantees. In particular, we introduce the notion of (α,ζ)-safe machine learning model. We propose a hypothesis testing procedure, based on the availability of a calibration set, to derive statistical guarantees providing that the probability of declaring that the adversarial (population) risk of a machine learning model is less than α (i.e. the model is safe), while the model is in fact unsafe (i.e. the model adversarial population risk is higher than α), is less than ζ. We also propose Bayesian optimization algorithms to determine efficiently whether a machine learning model is (α,ζ)-safe in the presence of an adversarial attack, along with statistical guarantees. We apply our framework to a range of machine learning models - including various sizes of vision Transformer (ViT) and ResNet models - impaired by a variety of adversarial attacks, such as PGDAttack, MomentumAttack, GenAttack and BanditAttack, to illustrate the operation of our approach. Importantly, we show that ViT's are generally more robust to adversarial attacks than ResNets, and large models are generally more robust than smaller models. Our approach goes beyond existing empirical adversarial risk-based certification guarantees. It formulates rigorous (and provable) performance guarantees that can be used to satisfy regulatory requirements mandating the use of state-of-the-art technical tools.
Given some video-query pairs with untrimmed videos and sentence queries, temporal sentence grounding (TSG) aims to locate query-relevant segments in these videos. Although previous respectable TSG methods have achieved remarkable success, they train each video-query pair separately and ignore the relationship between different pairs. To this end, in this paper, we pose a brand-new setting: Multi-Pair TSG, which aims to co-train these pairs. We propose a novel video-query co-training approach, Multi-Thread Knowledge Transfer Network, to locate a variety of video-query pairs effectively and efficiently. Firstly, we mine the spatial and temporal semantics across different queries to cooperate with each other. To learn intra- and inter-modal representations simultaneously, we design a cross-modal contrast module to explore the semantic consistency by a self-supervised strategy. To fully align visual and textual representations between different pairs, we design a prototype alignment strategy to 1) match object prototypes and phrase prototypes for spatial alignment, and 2) align activity prototypes and sentence prototypes for temporal alignment. Finally, we develop an adaptive negative selection module to adaptively generate a threshold for cross-modal matching. Extensive experiments show the effectiveness and efficiency of our proposed method.
Event Cameras offer appealing advantages, including power efficiency and ultra-low latency, driving forward advancements in edge applications. In order to leverage mature frame-based algorithms, most approaches typically compute dense, image-like representations from sparse, asynchronous events. However, they are often unable to capture comprehensive information or are computationally intensive, which hinders the edge deployment of event-based vision. Meanwhile, pillar-based paradigms have been proven to be efficient and well established for dense representations of sparse data. Hence, from a novel pillar-based perspective, we present EventPillars, an efficient, comprehensive framework for dense event representations. To summarize, it (i) incorporates the Temporal Event Range to describe an intact temporal distribution, (ii) Activates the Event Polarities to explicitly record the scene dynamics, (iii) enhances the target awareness by a spatial attention prior from Normalized Event Density, (iv) can be plug-and-played into different downstream tasks. Extensive experiments show that our EventPillars records a new state-of-the-art precision on object recognition and detection datasets with surprisingly 9.2× and 4.5× lower computation and storage consumption. This brings a new insight into dense event representations and is promising to boost the edge deployment of event-based vision.