ACM MM 2024 论文列表

Current co-speech motion generation approaches usually focus on upper body gestures following speech contents only, while lacking supporting the elaborate control of synergistic full-body motion based on text prompts, such as talking while walking. The major challenges lie in 1) the existing speech-to-motion datasets only involve highly limited full-body motions, making a wide range of common human activities out of training distribution; 2) these datasets also lack annotated user prompts. To address these challenges, we propose SynTalker, which utilizes the off-the-shelf text-to-motion dataset as an auxiliary for supplementing the missing full-body motion and prompts. The core technical contributions are two-fold. One is the multi-stage training process which obtains an aligned embedding space of motion, speech, and prompts despite the significant distributional mismatch in motion between speech-to-motion and text-to-motion datasets. Another is the diffusion-based conditional inference process, which utilizes the separate-then-combine strategy to realize fine-grained control of local body parts. Extensive experiments are conducted to verify that our approach supports precise and flexible control of synergistic full-body motion generation based on both speeches and user prompts, which is beyond the ability of existing approaches.

Some video traffic carries harmful content, such as hate speech and child abuse, primarily encrypted and transmitted through Dynamic Adaptive Streaming over HTTP (DASH). Promptly identifying and intercepting traffic of harmful videos is crucial in network regulation. However, QUIC is becoming another DASH transport protocol in addition to TCP. On the other hand, complex network environments and diverse playback modes lead to significant distortions in traffic. The issues above have not been effectively addressed. This paper proposes a real-time identification method for DASH encrypted video traffic with distortion, named Zenith. We extract stable video segment sequences under various itags as video fingerprints to tackle resolution changes and propose a method of traffic fingerprint extraction under QUIC and VPN. Subsequently, simulating the sequence matching problem as a natural language problem, we propose Traffic Language Model (TLM), which can effectively address video data loss and retransmission. Finally, we propose a frequency dictionary to accelerate Zenith's speed further. Zenith significantly improves accuracy and speed compared to other SOTA methods in various complex scenarios, especially in QUIC, VPN, automatic resolution, and low bandwidth. Zenith requires traffic for just half a minute of video content to achieve precise identification, demonstrating its real-time effectiveness.

As VR devices become increasingly prevalent, live 360-degree video has surged in popularity. However, current live 360-degree video systems heavily rely on uplink bandwidth to deliver high-quality live videos. Recent advancements in neural-enhanced streaming offer a promising solution to this limitation by leveraging server-side computation to conserve bandwidth. Nevertheless, these methods have primarily concentrated on neural enhancement within a single domain (either spatial or temporal), which may not adeptly adapt to diverse video scenarios and fluctuating bandwidth conditions. In this paper, we propose Lumos, a novel spatial-temporal integrated neural-enhanced live 360-degree video streaming system. To accommodate varied video scenarios, we devise a real-time Neural-enhanced Quality Prediction (NQP) model to predict the neural-enhanced quality for different video contents. To cope with varying bandwidth conditions, we design a Content-aware Bitrate Allocator, which dynamically allocates bitrates and selects an appropriate neural enhancement configuration based on the current bandwidth. Moreover, Lumos employs online learning to improve prediction performance and adjust resource utilization to optimize user quality of experience (QoE). Experimental results demonstrate that Lumos surpasses state-of-the-art neural-enhanced systems with an improvement of up to 0.022 in terms of SSIM, translating to an 8.2%-8.5% enhancement in QoE for live stream viewers.

In this paper, we tackle the new task of video-based Activated Muscle Group Estimation (AMGE) aiming at identifying active muscle regions during physical activity in the wild.To this intent, we provide the MuscleMap dataset featuring >15𝐾 video clips with 135 different activities and 20 labeled muscle groups.This dataset opens the vistas to multiple video-based applications in sports and rehabilitation medicine under flexible environment constraints.The proposed MuscleMap dataset is constructed with YouTube videos, specifically targeting High-Intensity Interval Training (HIIT) physical exercise in the wild.To make the AMGE model applicable in real-life situations, it is crucial to ensure that the model can generalize well to numerous types of physical activities not present during training and involving new combinations of activated muscles.To achieve this, our benchmark also covers an evaluation setting where the model is exposed to activity types excluded from the training set.Our experiments reveal that the generalizability of existing architectures adapted for the AMGE task remains a challenge.Therefore, we also propose a new approach, TransM 3 E, which employs a multi-modality feature fusion mechanism between both the video transformer model and the skeleton-based graph convolution model with novel cross-modal knowledge distillation executed on multiclassification tokens.The proposed method surpasses all popular video classification models when dealing with both, previously seen and new types of physical activities.The database and code can be found at https://github.com/KPeng9510/MuscleMap.

Classifying videos differs from that of images in the need to capture the information on what has happened, instead of what is in the frames. Conventional methods typically follow the data-driven approach, which uses transformer-based attention models to extract and aggregate the features of video frames as the representation of the entire video. However, this approach tends to extract the object information of frames and may face difficulties in classifying the classes talking about events, such as "fixing bicycle". To address this issue, This paper presents an Event-level Causal Representation Learning (ECRL) model for the spatio-temporal modeling of both the in-frame object interactions and their cross-frame temporal correlations. Specifically, ECRL first employs a Frame-to-Video Causal Modeling (F2VCM) module, which simultaneously builds the in-frame causal graph with the background and foreground information and models their cross-frame correlations to construct a video-level causal graph. Subsequently, a Causality-aware Event-level Representation Inference (CERI) module is introduced to eliminate the spurious correlations in contexts and objects via the back- and front-door interventions, respectively. The former involves visual context de-biasing to filter out background confounders, while the latter employs global-local causal attention to capture event-level visual information. Experimental results on two benchmarking datasets verified that ECRL may better capture the cross-frame correlations to describe videos in event-level features. The source codes have been released at https://github.com/wyqcrystal/ECRL.

By leveraging multi-view inputs to synthesize novel-view images, Neural Radiance Fields (NeRF) have emerged as a prominent technique in the realm of 3D object reconstruction. However, existing methods primarily focus on global scene reconstruction using large datasets, which necessitate substantial computational resources and impose high-quality requirements on input images. Nevertheless, in practical applications, users prioritize the 3D reconstruction results of on-demand specific object (OSO) based on their individual demands . Furthermore, the collected images transmitted through high-interference wireless environment (HIWE) leads to negatively impact the accuracy of NeRF reconstruction, thereby limiting its scalability. In this paper, we propose a novel on-demand Semantic Neural Radiance Fields (OSNeRF) scheme, which offers fast and robust 3D object reconstruction for diverse tasks. Within OSNeRF, semantic encoder is employed to extract core semantic features of OSOs from the collected scene images, semantic decoder is utilized to facilitate robust image recovery under HIWE conditions, lightweight renderer is employed for fast and efficient object reconstruction. Moreover, a semantic control unit (SCU) is introduced to guide above components, thereby enhancing the efficiency of reconstruction. Demonstrative experiments demonstrate that the proposed OSNeRF enables fast and robust object reconstruction in HIWE, surpassing the performance of state-of-the-art (SOTA) methods in terms of reconstruction quality.

Arbitrary style transfer aims to render artistic features from a style reference onto an image while retaining its original content. Previous methods either focus on learning the holistic style from a specific artist or extracting instance features from a single artwork. However, they often fail to apply style elements uniformly across the entire image and lack adaptation to the style of different artworks. To solve these issues, our key insight is that the art genre has better generality and adaptability than the overall features of the artist. To this end, we propose a Dual-head Genre-instance Transformer (DGiT) framework to simultaneously capture the genre and instance features for arbitrary style transfer. To the best of our knowledge, this is the first work to integrate the genre features and instance features to generate a high-quality stylized image. Moreover, we design two contrastive losses to enhance the capability of the network to capture two style features. Our approach ensures the uniform distribution of the overall style across the stylized image while enhancing the details of textures and strokes in local regions. Qualitative and quantitative evaluations demonstrate that our approach exhibits superior visual quality and efficiency.

Annotating data for sensitive labels (e.g., disease, smoking) poses a potential threats to individual privacy in many real-world scenarios. To cope with this problem, we propose a novel setting to protect privacy of each instance, namely learning from concealed labels for multi-class classification. Concealed labels prevent sensitive labels from appearing in the label set during the label collection stage, which specifies none and some random sampled insensitive labels as concealed labels set to annotate sensitive data. In this paper, an unbiased estimator can be established from concealed data under mild assumptions, and the learned multi-class classifier can not only classify the instance from insensitive labels accurately but also recognize the instance from the sensitive labels. Moreover, we bound the estimation error and show that the multi-class classifier achieves the optimal parametric convergence rate. Experiments demonstrate the significance and effectiveness of the proposed method for concealed labels in synthetic and real-world datasets. Source code is available at https://github.com/WilsonMqz/CLF

Face super-resolution aims to reconstruct a high-resolution face image from a low-resolution face image. Previous methods typically employ an encoder-decoder structure to extract facial structural features, where the direct downsampling inevitably introduces distortions, especially to high-frequency features such as edges. To address this issue, we propose a wavelet-based feature enhancement network, which mitigates feature distortion by losslessly decomposing the input feature into high and low-frequency components using the wavelet transform and processing them separately. To improve the efficiency of facial feature extraction, a full domain Transformer is further proposed to enhance local, regional, and global facial features. Such designs allow our method to perform better without stacking many modules as previous methods did. Experiments show that our method effectively balances performance, model size, and speed. Code link: https://github.com/PRIS-CV/WFEN.

Diffusion models have shown impressive potential on talking head generation. While plausible appearance and talking effect are achieved, these methods still suffer from temporal, 3D or expression inconsistency due to the error accumulation and inherent limitation of single-image generation ability. In this paper, we propose ConsistentAvatar, a novel framework for fully consistent and high-fidelity talking avatar generation. Instead of directly employing multi-modal conditions to the diffusion process, our method learns to first model the temporal representation for stability between adjacent frames. Specifically, we propose a Temporally-Sensitive Detail (TSD) map containing high-frequency feature and contours that vary significantly along the time axis. Using a temporal consistent diffusion module, we learn to align TSD of the initial result to that of the video frame ground truth. The final avatar is generated by a fully consistent diffusion module, conditioned on the aligned TSD, rough head normal, and emotion prompt embedding. We find that the aligned TSD, which represents the temporal patterns, constrains the diffusion process to generate temporally stable talking head. Further, its reliable guidance complements the inaccuracy of other conditions, suppressing the accumulated error while improving the consistency on various aspects. Extensive experiments demonstrate that ConsistentAvatar outperforms the state-of-the-art methods on the generated appearance, 3D, expression and temporal consistency.

In recent years, immersive communication has emerged as a compelling alternative to traditional video communication methods. One prospective avenue for immersive communication involves augmenting the user's immersive experience through the transmission of three-dimensional (3D) talking heads (THs). However, transmitting 3D THs poses significant challenges due to its complex and voluminous nature, often leading to pronounced distortion and a compromised user experience. Addressing this challenge, we introduce the 3D Talking Heads Quality Assessment (THQA-3D) dataset, comprising 1,000 sets of distorted and 50 original TH mesh sequences (MSs), to facilitate quality assessment in 3D TH transmission. A subjective experiment, characterized by a novel interactive approach, is conducted with recruited participants to assess the quality of MSs in THQA-3D dataset. Leveraging this dataset, we also propose a multimodal Quality-of-Experience (QoE) method incorporating a Large Quality Model (LQM). This method involves frontal projection of MSs and subsequent rendering into videos, with quality assessment facilitated by the LQM and a variable-length video memory filter (VVMF). Additionally, tone-lip coherence and silence detection techniques are employed to characterize audio-visual coherence in 3D MS streams. Experimental evaluation demonstrates the proposed method's superiority, achieving state-of-the-art performance on the THQA-3D dataset and competitiveness on other QoE datasets. Both the THQA-3D dataset and the QoE model have been publicly released at https://github.com/zyj-2000/THQA-3D

Deep networks have shown impressive performance in the image restoration tasks, such as image colorization. However, we find that previous approaches rely on the digital representation from single color model with a specific mapping function, a.k.a., color space, during the colorization pipeline. In this paper, we first investigate the modeling of different color spaces, and find each of them exhibiting distinctive characteristics with unique distribution of colors. The complementarity among multiple color spaces leads to benefits for the image colorization task. We present MultiColor, a new learning-based approach to automatically colorize grayscale images that combines clues from multiple color spaces. Specifically, we employ a set of dedicated colorization modules for individual color space. Within each module, a transformer decoder is first employed to refine color query embeddings and then a color mapper produces color channel prediction using the embeddings and semantic features. With these predicted color channels representing various color spaces, a complementary network is designed to exploit the complementarity and generate pleasing and reasonable colorized images. We conduct extensive experiments on real-world datasets, and the results demonstrate superior performance over the state-of-the-arts.

In this work, we focus on exploring explicit fine-grained control of generative facial image editing, all while generating faithful facial appearances and consistent semantic details, which however, is quite challenging and has not been extensively explored, especially under an one-shot scenario. We identify the key challenge as the exploration of disentangled conditional control between high-level semantics and explicit parameters (e.g., 3DMM) in the generation process, and accordingly propose a novel diffusion-based editing framework, named DisControlFace. Specifically, we leverage a Diffusion Autoencoder (Diff-AE) as the semantic reconstruction backbone. To enable explicit face editing, we construct an Exp-FaceNet that is compatible with Diff-AE to generate spatial-wise explicit control conditions based on estimated 3DMM parameters. Different from current diffusion-based editing methods that train the whole conditional generative model from scratch, we freeze the pre-trained weights of the Diff-AE to maintain its semantically deterministic conditioning capability and accordingly propose a random semantic masking (RSM) strategy to effectively achieve an independent training of Exp-FaceNet. This setting endows the model with disentangled face control meanwhile reducing semantic information shift in editing. Our model can be trained using 2D in-the-wild portrait images without requiring 3D or video data and perform robust editing on any new facial image through a simple one-shot fine-tuning. Comprehensive experiments demonstrate that DisControlFace can generate realistic facial images with better editing accuracy and identity preservation over state-of-the-art methods. Project page: https://discontrolface.github.io/

Semantic edge detection (SED) is pivotal for the precise demarcation of object boundaries, yet it faces ongoing challenges due to the prevalence of low-quality labels in current methods. In this paper, we present a novel solution to bolster SED through the encoding of both language and image data. Distinct from antecedent language-driven techniques, which predominantly utilize static elements such as dataset labels, our method taps into the dynamic language content that details the objects in each image and their interrelations. By encoding this varied input, we generate integrated features that utilize semantic insights to refine the high-level image features and the ultimate mask representations. This advancement improves the quality of these features and elevates SED performance. Experimental evaluation on benchmark datasets, including SBD and Cityscape, showcases the efficacy of our method, achieving leading ODS F-scores of 79.0 and 76.0, respectively. Our approach signifies a notable advancement in SED technology by seamlessly integrating multimodal textual information, embracing both static and dynamic aspects.

Recently, there has been significant progress in leveraging human feedback to enhance diffusion-based image generation, garnering considerable interest and attention. However, existing methods fail to achieve a fine-grained performance boost for the following challenges: i) insufficient amount of fine-grained feedback data; ii) lack of effective fine-grained feedback learning framework; To tackle these challenges, we present TreeReward to facilitate the fine-grained feedback optimization for diffusion models. Specifically, to address the limitation of the fine-grained feedback data, we first design a novel "AI + Expert" feedback data construction pipeline, yielding about 2.2M high-quality feedback dataset encompassing six fine-grained dimensions at a relatively low cost. Built upon this dataset, we introduce a tree-structure reward model to exploit the fine-grained feedback data efficiently and provide tailored optimization during feedback learning. We validate the feedback learning performance of our method across different fine-grained dimensions and various downstream tasks. Extensive experiments on both Stable Diffusion v1.5 (SD1.5) and Stable Diffusion XL (SDXL) demonstrate the effectiveness of our method in enhancing the general and fine-grained generation and downstream tasks generalization.

Text-to-3D content creation has recently received much attention, especially with the prevalence of 3D Gaussians Splatting (3D GS). In general, GS-based methods comprise two key stages: initialization and rendering optimization. To achieve initialization, existing works directly apply random sphere initialization or 3D diffusion models, e.g., Point-E, to derive the initial shapes. However, such strategies suffer from two critical yet challenging problems: 1) the final shapes are still similar to the initial ones even after training; 2) shapes can be produced only from simple texts, e.g., "a dog", not for lexically richer (or harder) texts, e.g., "a dog is sitting on the top of the airplane". To address these problems, this paper proposes a novel general framework to boost the 3D GS Initialization for text-to-3D generation upon the lexical richness. Our key idea is to aggregate 3D Gaussians into spatially uniform voxels to represent complex shapes while enabling the spatial interaction among the 3D Gaussians and semantic interaction between Gaussians and texts. Specifically, we first construct a voxelized representation, where each voxel holds a 3D Gaussian with its position, scale, and rotation fixed while setting opacity as the sole factor to determine a position's occupancy. We then design an initialization network mainly consisting of two novel components: 1) Global Information Perception (GIP) block and 2) Gaussians-Text Fusion (GTF) block. Such a design enables each 3D Gaussian to assimilate the spatial information from other areas and semantic information from texts. Extensive experiments show the superiority of our framework of high-quality 3D GS initialization against the existing methods, e.g., Shap-E, by taking lexically simple, medium, and hard texts. Also, our framework can be seamlessly plugged into state-of-the-art training frameworks, e.g., LucidDreamer for semantically consistent text-to-3D generation. The project code is available at https://vlislab22.github.io/DreamInit/.

Knowledge distillation has become widely recognized for its ability to transfer knowledge from a large teacher network to a compact and more streamlined student network. Traditional knowledge distillation methods primarily follow a teacher-oriented paradigm that imposes the task of learning the teacher's complex knowledge onto the student network. However, significant disparities in model capacity and architectural design hinder the student's comprehension of the complex knowledge imparted by the teacher, resulting in sub-optimal performance. This paper introduces a novel perspective emphasizing student-oriented and refining the teacher's knowledge to better align with the student's needs, thereby improving knowledge transfer effectiveness. Specifically, we present the Student-Oriented Knowledge Distillation (SoKD), which incorporates a learnable feature augmentation strategy during training to refine the teacher's knowledge of the student dynamically. Furthermore, we deploy the Distinctive Area Detection Module (DAM) to identify areas of mutual interest between the teacher and student, concentrating knowledge transfer within these critical areas to avoid transferring irrelevant information. This customized module ensures a more focused and effective knowledge distillation process. Our approach, functioning as a plug-in, could be integrated with various knowledge distillation methods. Extensive experimental results demonstrate the efficacy and generalizability of our method.

In the last two years, Artificial Intelligence Generated Content (AIGC) has received significant attention, leading to an anecdotal rise in the amount of AIGC being shared via social media platforms. The impact of AIGC and its implications are of key importance to social platforms, e.g., regarding the implementation of policies, community formation, and algorithmic design. Yet, to date, we know little about how the arrival of AIGC has impacted the social media ecosystem. To fill this gap, we present a comprehensive study of Pixiv, an online community for artists who wish to share and receive feedback on their illustrations. Pixiv hosts over 100 million artistic submissions and receives more than 1 billion page views per month (as of 2023). Importantly, it allows both human and AI generated content to be uploaded. Exploiting this, we perform the first analysis of the impact that AIGC has had on the social media ecosystem, through the lens of Pixiv. Based on a dataset of 15.2 million posts (including 2.4 million AI-generated images), we measure the impact of AIGC on the Pixiv community, as well as the differences between AIGC and human-generated content in terms of content creation and consumption patterns. Our results offer key insight to how AIGC is changing the dynamics of social media platforms like Pixiv.

With global occurrences of crowd crushes and stampedes, dense crowd simulation has been drawing great attention. In this research, our goal is to simulate dense crowd motions under six classic motion patterns, more specifically, to generate subsequent motions of dense crowds from the given initial states. Since dense crowds share similarities with fluids, such as continuity and fluidity, one common approach for dense crowd simulation is to construct hydrodynamics-based models, which consider dense crowds as fluids, guide crowd motions with Navier-Stokes equations, and conduct dense crowd simulation by solving governing equations. Despite the proposal of these models, dense crowd simulation faces multiple challenges, including the difficulty of directly solving Navier-Stokes equations due to their nonlinear nature, the ignorance of distinctive crowd characteristics which fluids lack, and the gaps in the evaluation and validation of crowd simulation models. To address the above challenges, we build a hydrodynamic model, which captures the crowd physical properties (continuity, fluidity, etc.) with Navier-Stokes equations and reflects the crowd social properties (sociality, personality, etc.) with operators that describe crowd interactions and crowd-environment interactions. To tackle the computational problem, we propose to solve the governing equation based on Navier-Stokes equations using neural networks, and introduce the Hydrodynamics-Informed Neural Network (HINN) which preserves the structure of the governing equation in its network architecture. To facilitate the evaluation, we construct a new dense crowd motion video dataset called Dense Crowd Flow Dataset (DCFD), containing six classic motion patterns (line, curve, circle, cross, cluster and scatter) and 457 video clips, which can serve as the groundtruths for various objective metrics. Numerous experiments are conducted using HINN to simulate dense crowd motions under six motion patterns with video clips from DCFD. Objective evaluation metrics that concerns authenticity, fidelity and diversity demonstrate the superior performance of our model in dense crowd simulation compared to other simulation models. Our code and dataset are available at https://github.com/shanshan-zys/HINN.