Is the Text to Motion model robust? Recent advancements in Text to Motion models primarily stem from more accurate predictions of specific actions. However, the text modality typically relies solely on pre-trained Contrastive Language-Image Pretraining (CLIP) models. Our research has uncovered a significant issue with the text-to-motion model: its predictions often exhibit inconsistent outputs, resulting in vastly different or even incorrect poses when presented with semantically similar or identical text inputs. In this paper, we undertake an analysis to elucidate the underlying causes of this instability, establishing a clear link between the unpredictability of model outputs and the erratic attention patterns of the text encoder module. Consequently, we introduce a formal framework aimed at addressing this issue, which we term the Sta ble Text-to-Motion Framework (SATO). SATO consists of three modules, each dedicated to stable attention, stable prediction, and maintaining a balance between accuracy and robustness trade-off. We present a methodology for constructing an SATO that satisfies the stability of attention and prediction. To verify the stability of the model, we introduced a new textual synonym perturbation dataset based on HumanML3D and KIT-ML. Results show that SATO is significantly more stable against synonyms and other slight perturbations while keeping its high accuracy performance. Codes and models are released at https://github.com/sato-team/Stable-Text-to-Motion-Framework.
论文检索
输入标题、作者或关键词，从 2,072 篇学术成果中精准定位
Despite having tremendous progress in image-to-3D generation, existing methods still struggle to produce multi-view consistent images with high-resolution textures in detail, especially in the paradigm of 2D diffusion that lacks 3D awareness. In this work, we present High-resolution Image-to-3D model (Hi3D), a new video diffusion based paradigm that redefines a single image to multi-view images as 3D-aware sequential image generation (i.e., orbital video generation). This methodology delves into the underlying temporal consistency knowledge in video diffusion model that generalizes well to geometry consistency across multiple views in 3D generation. Technically, Hi3D first empowers the pre-trained video diffusion model with 3D-aware prior (camera pose condition), yielding multi-view images with low-resolution texture details. A 3D-aware video-to-video refiner is learnt to further scale up the multi-view images with high-resolution texture details. Such high-resolution multi-view images are further augmented with novel views through 3D Gaussian Splatting, which are finally leveraged to obtain high-fidelity meshes via 3D reconstruction. Extensive experiments on both novel view synthesis and single view reconstruction demonstrate that our Hi3D manages to produce superior multi-view consistency images with highly-detailed textures. Source code and data are available at https://github.com/yanghb22-fdu/Hi3D-Official.
Visual prompting is an efficient methodology for finetuning pretrained visual models by introducing a small number of learnable parameters while keeping the backbone frozen. However, most existing visual prompting methods learn a shared prompt for all samples, making it challenging to grasp distinct characteristics among diverse samples, thereby limiting the model's performance. While other methods partially address this issue through sample clustering and learning multiple prompts, they still struggle to capture nuanced differences among instances and incur significant parameter overhead. Therefore, to comprehensively and efficiently leverage discriminative characteristics of individual instances, we propose an Instance Visual Prompting method, called InsVP. Initially, the instance image prompt is introduced to extract both crucial and nuanced discriminative information from the original image itself and is overlaid onto the input image. Furthermore, the instance feature prompt is designed to capture both commonalities and characteristics among individual instances, fed into the model's intermediate layers to facilitate feature extraction. Consequently, the instance image and feature prompts complement each other, enhancing the adaptation ability of pretrained models to extract discriminative features from individual instances. Extensive experiments on various large-scale benchmarks show that our InsVP achieves superior performance exceeding the state-of-the-art methods at a lower parameter cost. The code is available at https://github.com/zhoujiahuan1991/MM2024-InsVP .
The domain of image restoration encompasses a wide array of highly effective models (e.g., SwinIR, CODE, DnCNN), each exhibiting distinct advantages in either efficiency or performance. Selecting and deploying these models necessitate careful consideration of resource limitations. While some studies have explored dynamic restoration through the integration of an auxiliary network within a unified framework, these approaches often fall short in practical applications due to the complexities involved in training, retraining, and hyperparameter adjustment, as well as limitations as being totally controlled by auxiliary network and biased by training data. To address these challenges, we introduce FlexIR: a flexible and manipulable framework for image restoration. FlexIR is distinguished by three components: a meticulously designed hierarchical branch network enabling dynamic output, an innovative progressive self-distillation process, and a channel-wise evaluation method to enhance knowledge distillation efficiency. Additionally, we propose two novel inference methodologies to fully leverage FlexIR, catering to diverse user needs and deployment contexts. Through this framework, FlexIR achieves unparalleled performance across all branches, allowing users to navigate the trade-offs between quality, cost, and efficiency during the inference phase. Crucially, FlexIR employs a dynamic mechanism powered by a non-learning metric independent of training data, ensuring that FlexIR is entirely under the direct control of the user. Comprehensive experimental evaluations validate FlexIR's flexibility, manipulability, and cost-effectiveness, showcasing its potential for straightforward adjustments and quick adaptations across a range of scenarios.
The rise of social media and the exponential growth of multimodal communication necessitates advanced techniques for Multimodal Information Extraction (MIE). However, existing methodologies primarily rely on direct Image-Text interactions, a paradigm that often faces significant challenges due to semantic and modality gaps between images and text. In this paper, we introduce a new paradigm of Image-Context-Text interaction, where large multimodal models (LMMs) are utilized to generate descriptive textual context to bridge these gaps. In line with this paradigm, we propose a novel Shapley Value-based Contrastive Alignment (Shap-CA) method, which aligns both context-text and context-image pairs. Shap-CA initially applies the Shapley value concept from cooperative game theory to assess the individual contribution of each element in the set of contexts, texts and images towards total semantic and modality overlaps. Following this quantitative evaluation, a contrastive learning strategy is employed to enhance the interactive contribution within context-text/image pairs, while minimizing the influence across these pairs. Furthermore, we design an adaptive fusion module for selective cross-modal fusion. Extensive experiments across four MIE datasets demonstrate that our method significantly outperforms existing state-of-the-art methods.
In Large Language Models (LLMs), text generation that involves knowledge representation is often fraught with the risk of "hallucinations'', where models confidently produce erroneous or fabricated content. These inaccuracies often stem from intrinsic biases in the pre-training stage or from the incorporation of human preference biases during the fine-tuning process. To mitigate these issues, we take inspiration from Goldman's causal theory of knowledge, which asserts that knowledge is not merely about having a true belief but also involves a causal connection between the belief and the truth of the proposition. We instantiate this theory within the context of Knowledge Question Answering (KQA) by constructing a causal graph that delineates the pathways between the candidate knowledge and belief. Through the application of the do-calculus rules from structural causal models, we devise an unbiased estimation framework based on this causal graph, thereby establishing a methodology for knowledge modeling grounded in causal inference. The resulting CORE framework (short for "Causal knOwledge REasoning'') is comprised of four essential components: question answering, causal reasoning, belief scoring, and refinement. Together, they synergistically improve the KQA system by fostering faithful reasoning and introspection. Extensive experiments are conducted on ScienceQA and HotpotQA datasets, which demonstrate the effectiveness and rationality of the CORE framework.
Multimodal Question Answering (MMQA) is crucial as it enables comprehensive understanding and accurate responses by integrating insights from diverse data representations such as tables, charts, and text. Most existing researches in MMQA only focus on two modalities such as image-text QA, table-text QA and chart-text QA, and there remains a notable scarcity in studies that investigate the joint analysis of text, tables, and charts. In this paper, we present CT2C-QA, a pioneering Chinese reasoning-based QA dataset that includes an extensive collection of text, tables, and charts, meticulously compiled from 200 selectively sourced webpages. Our dataset simulates real webpages and serves as a great test for the capability of the model to analyze and reason with multimodal data, because the answer to a question could appear in various modalities, or even potentially not exist at all. Additionally, we present AED (Allocating, Expert and Decision), a multi-agent system implemented through collaborative deployment, information interaction, and collective decision-making among different agents. Specifically, the Assignment Agent is in charge of selecting and activating expert agents, including those proficient in text, tables, and charts. The Decision Agent bears the responsibility of delivering the final verdict, drawing upon the analytical insights provided by these expert agents. We execute a comprehensive analysis, comparing AED with various state-of-the-art models in MMQA, including GPT-4. The experimental outcomes demonstrate that current methodologies, including GPT-4, are yet to meet the benchmarks set by our dataset.
In practical object detection scenarios, distributed data and stringent privacy protections significantly limit the feasibility of traditional centralized training methods. Federated learning (FL) emerges as a promising solution to this dilemma. Nonetheless, the issue of data heterogeneity introduces distinct challenges to federated object detection, evident in diminished object perception, classification and localization abilities. In response, we introduce a task-driven federated learning methodology, dubbed Adaptive Hierarchical Aggregation (FedAHA), tailored to overcome these obstacles. Our algorithm unfolds in two strategic phases from shallow-to-deep layers: (1) Structure-aware Aggregation (SAA) aligns feature extractors during the aggregation phase, thus bolstering the global model's object perception capabilities; (2) Convex Semantic Calibration (CSC) leverages convex function theory to average semantic features instead of model parameters, enhancing the global model's classification and localization precision. We demonstrate experimentally and theoretically the effectiveness of the proposed two modules respectively. Our method consistently outperforming the state-of-the-art methods across multiple valuable application scenarios from 2.26% to 7.61%. Moreover, we build a real FL system using Raspberry Pis to demonstrate that our approach achieves a good trade-off between performance and efficiency.
Text-to-image (T2I) generation is a pivotal and core interest within the realm of AI content generation. Amid the swift advancements of both open-source (such as Stable Diffusion) and proprietary (for example, DALLE, MidJourney) T2I models, there is a notable absence of a comprehensive and robust quantitative framework for evaluating their output quality. Traditional methods of quality assessment overlook the textual prompts when judging images; meanwhile, the advent of large multi-modal models (LMMs) introduces the capability to incorporate text prompts in evaluations, yet the challenge of fine-tuning these models for precise T2I quality assessment remains unresolved. In our study, we introduce the T2I-Scorer, a novel two-stage training methodology aimed at fine-tuning LMMs for T2I evaluation. For the first stage, we collect 397K GPT-4V-labeled question-answer pairs related to T2I evaluation. Termed as T2I-ITD, the pseudo-labeled dataset is analyzed and examined by human, and used for instruction tuning to improve the LMM's low-level quality perception. The first stage model, T2I-Scorer-IT, has reached superior accuracy on T2I evaluation than all kinds of existing T2I metrics under zero-shot settings. For the second stage, we define an explicit multi-task training scheme to further align the LMM with human opinion scores, and the fine-tuned T2I-Scorer can reach state-of-the-art accuracy on both image quality and image-text alignment perspectives with significant improvements. We anticipate the proposed metrics can serve as a reliable metric to gauge the ability of T2I generation models in the future. We will make code, data, and weights publicly available.
Arondight: Red Teaming Large Vision Language Models with Auto-generated Multi-modal Jailbreak Prompts
PDF ↗Large Vision Language Models (VLMs) extend and enhance the perceptual abilities of Large Language Models (LLMs). Despite offering new possibilities for LLM applications, these advancements raise significant security and ethical concerns, particularly regarding the generation of harmful content. While LLMs have undergone extensive security evaluations with the aid of red teaming frameworks, VLMs currently lack a well-developed one. To fill this gap, we introduce Arondight, a standardized red team framework tailored specifically for VLMs. Arondight is dedicated to resolving issues related to the absence of visual modality and inadequate diversity encountered when transitioning existing red teaming methodologies from LLMs to VLMs. Our framework features an automated multi-modal jailbreak attack, wherein visual jailbreak prompts are produced by a red team VLM, and textual prompts are generated by a red team LLM guided by a reinforcement learning agent. To enhance the comprehensiveness of VLM security evaluation, we integrate entropy bonuses and novelty reward metrics. These elements incentivize the RL agent to guide the red team LLM in creating a wider array of diverse and previously unseen test cases. Our evaluation of ten cutting-edge VLMs exposes significant security vulnerabilities, particularly in generating toxic images and aligning multi-modal prompts. In particular, our Arondight achieves an average attack success rate of 84.5% on GPT-4 in all fourteen prohibited scenarios defined by OpenAI in terms of generating toxic text. For a clearer comparison, we also categorize existing VLMs based on their safety levels and provide corresponding reinforcement recommendations. Our multimodal prompt dataset and red team code will be released after ethics committee approval. CONTENT WARNING: THIS PAPER CONTAINS HARMFUL MODEL RESPONSES.
Recent advancements in the field of Diffusion Transformers have substantially improved the generation of high-quality 2D images, 3D videos, and 3D shapes. However, the effectiveness of the Transformer architecture in the domain of co-speech gesture generation remains relatively unexplored, as prior methodologies have predominantly employed the Convolutional Neural Network (CNNs) or simple a few transformer layers. In an attempt to bridge this research gap, we introduce a novel Masked Diffusion Transformer for co-speech gesture generation, referred to as MDT-A2G, which directly implements the denoising process on gesture sequences. To enhance the contextual reasoning capability of temporally aligned speech-driven gestures, we incorporate a novel Masked Diffusion Transformer. This model employs a mask modeling scheme specifically designed to strengthen temporal relation learning among sequence gestures, thereby expediting the learning process and leading to coherent and realistic motions. Apart from audio, Our MDT-A2G model also integrates multi-modal information, encompassing text, emotion, and identity. Furthermore, we propose an efficient inference strategy that diminishes the denoising computation by leveraging previously calculated results, thereby achieving a speedup with negligible performance degradation. Experimental results demonstrate that MDT-A2G excels in gesture generation, boasting a learning speed that is over 6× faster than traditional diffusion transformers and an inference speed that is 5.7× than the standard diffusion model. Our code is available at https://xiaofenmao.github.io/web-project/MDT-A2G/
With the growth of VR and AR industry, 3D reconstruction has become a more and more important topic in multimedia. Although 3D Gaussian Splatting achieves state-of-the-art in 3D Reconstruction, a large number of Gaussians are needed to fit a 3D scene due to the Gibbs Phenomenon. The pursuit of compressing 3D Gaussian Splatting and reducing memory overhead has long been a focal point. Embarking on this trajectory, our study delves into this domain, aiming to mitigate these challenges. Inspired by the tangram, a Chinese ancient puzzle, we introduce a novel methodology (Tangram-Splatting) that leverages shape priors to optimize 3D scene fitting. Central to our approach is a pioneering technique that diversifies Gaussian function types while preserving algorithmic efficiency. Through exhaustive experimentation, we demonstrate that our method achieves a remarkable average reduction of 62.4% in memory consumption used to store optimized parameters and decreases the training time by at least 10 minutes, with only marginal sacrifices in PSNR performance, typically under 0.3 dB, and our algorithm is even better on some datasets. This reduction in memory burden is of paramount significance for real-world applications, mitigating the substantial memory footprint and transmission burden traditionally associated with such algorithms. Our algorithm underscores the profound potential of Tangram-Splatting in advancing multimedia applications.
Recent advances in neural camera imaging pipelines have demonstrated notable progress. Nevertheless, the real-world imaging pipeline still faces challenges including the lack of joint optimization in system components, computational redundancies, and optical distortions such as lens shading. In light of this, we propose an end-to-end camera imaging pipeline (RealCamNet) to enhance real-world camera imaging performance. Our methodology diverges from conventional, fragmented multi-stage image signal processing towards end-to-end architecture. This architecture facilitates joint optimization across the full pipeline and the restoration of coordinate-biased distortions. RealCamNet is designed for high-quality conversion from RAW to RGB and compact image compression. Specifically, we deeply analyze coordinate-dependent optical distortions, e.g., vignetting and dark shading, and design a novel Coordinate-Aware Distortion Restoration (CADR) module to restore coordinate-biased distortions. Furthermore, we propose a Coordinate-Independent Mapping Compression (CIMC) module to implement tone mapping and redundant information compression. Existing datasets suffer from misalignment and overly idealized conditions, making them inadequate for training real-world imaging pipelines. Therefore, we collected a real-world imaging dataset. Experiment results show that RealCamNet achieves the best rate-distortion performance with lower inference latency.
In this paper, we propose a neural representation for videos that enables real-time quality-scalable decoding, called QS-NeRV. QS-NeRV comprises a Self-Learning Distribution Mapping Network (SDMN) and Extensible Enhancement Networks (EENs). Firstly, SDMN functions as the base layer (BL) for scalable video coding, focusing on encoding videos of lower quality. Within SDMN, we employ a methodology that minimizes the bitstream overhead to achieve efficient information exchange between the encoder and decoder instead of direct transmission. Specifically, we utilize an invertible network to map the multi-scale information obtained from the encoder to a specific distribution. Subsequently, during the decoding process, this information is recovered from a randomly sampled latent variable to assist the decoder in achieving improved reconstruction performance. Secondly, EENs serve as the enhancement layers (ELs) and are trained in an overfitting manner to obtain robust restoration capability. By integrating the fixed BL bitstream with the parameters of EEN as an extension pack, the decoder can produce higher-quality enhanced videos. Furthermore, the scalability of the method allows for adjusting the number of combined packs to accommodate diverse quality requirements. Experimental results demonstrate our proposed QS-NeRV outperforms the state-of-the-art real-time decoding INR-based methods on various datasets for video compression and interpolation tasks.
Amidst the prevailing trend of escalating demands for data and computational resources, the efficiency of data utilization emerges as a critical lever for enhancing the performance of deep learning models, especially in the realm of image restoration tasks. This investigation delves into the intricacies of data efficiency in the context of image restoration, with Gaussian image denoising serving as a case study. We postulate a strong correlation between the model's performance and the content information encapsulated in the training images. This hypothesis is rigorously tested through experiments conducted on synthetically blurred datasets. Building on this premise, we delve into the data efficiency within training datasets and introduce an effective and stabilized method for quantifying content information, thereby enabling the ranking of training images based on their influence. Our in-depth analysis sheds light on the impact of various subset selection strategies, informed by this ranking, on model performance. Furthermore, we examine the transferability of these efficient subsets across disparate network architectures. The findings underscore the potential to achieve comparable, if not superior, performance with a fraction of the data-highlighting instances where training IRCNN and Restormer models with only 3.89% and 2.30% of the data resulted in a negligible drop and, in some cases, a slight improvement in PSNR. This investigation offers valuable insights and methodologies to address data efficiency challenges in Gaussian denoising. Similarly, our method yields comparable conclusions in other restoration tasks. We believe this will be beneficial for future research.
Image aesthetics assessment (IAA) primarily examines image quality from a user-centric perspective and can be applied to guide various applications, including image capture, recommendation, and enhancement. The fundamental issue in IAA revolves around the quantification of image aesthetics. Existing methodologies rely on assigning a scalar (or a distribution) to represent aesthetic value based on conventional practices, which confines this scalar within a specific range and artificially labels it. However, conventional methods rarely incorporate research on interpretability, particularly lacking systematic responses to the following three fundamental questions: 1) Can aesthetic qualities be quantified? 2) What is the nature of quantifying aesthetics? 3) How can aesthetics be accurately quantified? In this paper, we present a law called "Special Relativity" of IAA (SR-IAA) that addresses the aforementioned core questions. We have developed a Multi-Attribute IAA Framework (MAINet), which serves as a preliminary validation for SR-IAA within the existing datasets and achieves state-of-the-art (SOTA) performance. Specifically, our metrics on multi-attribute assessment outperform the second-best performance by 8.06% (AADB), 1.67% (PARA), and 2.44% (SPAQ) in terms of SRCC. We anticipate that our research will offer innovative theoretical guidance to the IAA research community. All resources are available here.
3D Object Detection (3DOD) aims to accurately locate and identify 3D objects in point clouds, facing the challenge of balancing model performance with computational efficiency. Knowledge distillation emerges as a vital method for model compression in 3DOD, transferring knowledge from complex, larger models to smaller, efficient ones. However, the effectiveness of these methods is constrained by the intrinsic sparsity and structural complexity of point clouds. In this paper, we propose a novel methodology termed Joint Homophily and Heterophily Relational Knowledge Distillation (H2RKD) to distill robust relational knowledge in point clouds, thereby enhancing intra-object similarity and refining inter-object distinction. This unified strategy encompasses the integration of Collaborative Global Distillation (CGD) for distilling global relational knowledge across both distance and angular dimensions, and Separate Local Distillation (SLD) for a focused distillation of local relational dynamics. By seamlessly leveraging the relational dynamics within point clouds, the H2RKD facilitates a comprehensive knowledge transfer, significantly advancing 3D object detection capabilities. Extensive experiments on KITTI and unScenes datasets demonstrate the effectiveness of the proposed H2RKD.
We introduce HmPEAR, a novel dataset crafted for advancing research in 3D Human Pose Estimation (3D HPE) and Human Action Recognition (HAR), with a primary focus on outdoor environments. This dataset offers a synchronized collection of imagery, LiDAR point clouds, 3D human poses, and action categories. In total, the dataset encompasses over 300,000 frames collected from 10 distinct scenes and 25 diverse subjects. Among these, 250,000 frames of data contain 3D human pose annotations captured using an advanced motion capture system and further optimized for accuracy. Furthermore, the dataset annotates 40 types of daily human actions, resulting in over 6,000 action clips. Through extensive experimentation, we have demonstrated the quality of HmPEAR and highlighted the challenges it presents to current methodologies. Additionally, we propose baselines leveraging sequential images and point clouds for 3D HPE and HAR, which underscore the mutual reinforcement between them, highlighting the potential for cross-task synergies. The dataset is available at http://www.lidarhumanmotion.net/hmpear.
Multimodal Sentiment Analysis (MSA) has witnessed remarkable progress and gained increasing attention in recent decade. However, current MSA methodologies primarily rely on global representations extracted from different modalities, such as the mean of all token representations, to construct sophisticated fusion networks. These approaches often overlook the valuable details present in local representations, which consist of fused representations of consecutive several tokens. Additionally, the integration of multiple local representations, and the fusion of local and global information present significant challenges. To address these limitations, we propose the Global-Local Modal (GLoMo) Fusion framework. It comprises two essential components: (i) modality-specific mixture of experts layers that integrate diverse local representations within each modality, and (ii) a global-guided fusion module that effectively combines global and local representations. The former component leverages specialized expert networks to automatically select and integrate crucial local representations from each modality, while the latter ensures the preservation of global information during the fusion process. We evaluate GLoMo on various datasets, encompassing tasks in multimodal sentiment analysis, multimodal humor detection, and multimodal emotion recognition. Extensive experiments demonstrate that GLoMo outperforms existing state-of-the-art models, validating the effectiveness of our proposed framework. Our code is publicly available at https://github.com/YetZzzzzz/GLoMo.
Document Visual Question Answering (DVQA) involves responding to queries based on the contents of document images. Existing works are confined to locating information within a single page and lack support for cross-page question-and-answer interactions. Furthermore, the token length limitation on model inputs can lead to the truncation of answer-relevant segments. In this study, we present CREAM, an innovative methodology that focuses on high-performance retrieval and integrates relevant multimodal document information to effectively address this critical issue. To overcome the limitations of current text embedding similarity methods, we first employ a coarse-to-fine retrieval and ranking approach. The coarse phase calculates the similarity between the query and text chunk embeddings, while the fine phase involves multiple rounds of grouping and ordering with a large language model to identify the text chunks most relevant to the query. Subsequently, integrating an attention pooling mechanism for multi-page document images into the vision encoder allows us to effectively merge the visual information of multi-page documents, enabling the multimodal large language model (MLLM) to simultaneously process both single-page and multi-page documents. Finally, we apply various parameter-efficient tuning methods to enhance document visual question-answering performance. Experiments demonstrate that our approach secures state-of-the-art results across various document datasets.