ACM MM 2024 论文列表

Domain shift significantly hinders crowd counting performance in unseen domains. Domain adaptation methods tackle this issue using target domain images but falter when acquiring these images is difficult. Moreover, they demand additional training time for fine-tuning. To address this issue, we propose an Uncertainty-Guided Style Diversity Augmentation (UGSDA) method, enabling the models to be trained solely on the source domain and directly generalized to various target domains. It is achieved by generating sufficiently diverse and realistic samples during the training process. Specifically, our UGSDA method incorporates three tailor-designed components: the Global Styling Elements Extraction (GSEE) module, the Local Uncertainty Perturbations (LUP) module, and the Density Distribution Consistency (DDC) loss. The GSEE extracts global style elements from the feature space of the whole source domain. The LUP aims to obtain uncertainty perturbations from the batch-level input to form style distributions beyond the source domain, which used to generate diversified stylized samples together with global style elements. To regulate the extent of perturbations, the DDC loss imposes constraints between the source samples and the stylized samples, ensuring the stylized samples maintain a higher degree of realism and reliability. Comprehensive experiments validate the superiority of our approach, demonstrating its strong generalization capabilities across various datasets and models. Code is available at https://github.com/gcding/UGSDA-pytorch.

Speaker extraction aims to selectively extract the target speaker from the multi-talker environment under the guidance of auxiliary reference. Recent studies have shown that the attended speaker's information can be decoded by the auditory attention decoding from the listener's brain activity. However, how to more effectively utilize the common information about the target speaker contained in both electroencephalography (EEG) and speech is still an unresolved problem. In this paper, we propose a multi-scale fusion network (MSFNet) for brain-controlled speaker extraction, which utilizes the EEG recorded from the listener to extract the target speech. In order to make full use of the speech information, the mixed speech is encoded with multiple time scales so that the multi-scale embeddings are acquired. In addition, to effectively extract the non-Euclidean data of EEG, the graph convolutional networks are used as the EEG encoder. Finally, these multi-scale embeddings are separately fused with the EEG features. To facilitate research related to auditory attention decoding and further validate the effectiveness of the proposed method, we also construct the AVED dataset, a new EEG-Audio dataset. Experimental results on both the public Cocktail Party dataset and the newly proposed AVED dataset in this paper show that our MSFNet model significantly outperforms the state-of-the-art method in certain objective evaluation metrics.

Mountains of researches center around the Remote Sensing Image-Text Retrieval (RSITR), aiming at retrieving the corresponding targets based on the given query. Among them, the transfer of Foundation Models (FMs), such as CLIP, to remote sensing domain shows promising results. However, existing FM-based approaches neglect the negative impact of weakly correlated sample pairs and the key distinctions among remote sensing texts, leading to biased and superficial exploration of sample pairs. To address these challenges, we propose a novel Eliminate Before Align strategy with Keyword Explicit Reasoning framework (EBAKER) for RSITR. Specifically, we devise an innovative Eliminate Before Align (EBA) strategy to filter out the weakly correlated sample pairs to mitigate their deviations from optimal embedding space during alignment. Moreover, we introduce a Keyword Explicit Reasoning (KER) module to facilitate the positive role of subtle key concept differences. Without bells and whistles, our method achieves a one-step transformation from FM to RSITR task, obviating the necessity for extra pretraining on remote sensing data. Extensive experiments on three popular benchmark datasets validate that our proposed EBAKER method outperform the state-of-the-art methods with fewer training data. Our source code will be released soon.

Research on Multi-modal Large Language Models (MLLMs) towards the multi-image cross-modal instruction has received increasing attention and made significant progress, particularly in scenarios involving closely resembling images ( e.g., change captioning). Existing MLLMs typically follow a two-step process in their pipelines: first, extracting visual tokens independently for each input image, and then aligning these visual tokens from different images with the Large Language Model (LLM) in its textual feature space. However, the independent extraction of visual tokens for each image may result in different semantics being prioritized for different images in the first step, leading to a lack of preservation of linking information among images for subsequent LLM analysis. This issue becomes more serious in scenarios where significant variations exist among the images (e.g., visual storytelling). To address this challenge, we introduce Semantic Alignment for Multi-modal large language models (SAM). By involving the bidirectional semantic guidance between different images in the visual-token extraction process, SAM aims to enhance the preservation of linking information for coherent analysis and align the semantics of different images before feeding them into LLM. As the test bed, we propose a large-scale dataset named MmLINK consisting of 69K samples. Different from most existing datasets for MLLMs fine-tuning, our MmLINK dataset comprises multi-modal instructions with significantly diverse images. Extensive experiments on the group captioning task and the storytelling task prove the effectiveness of our SAM model, surpassing the state-of-the-art methods by a large margin (+37% for group captioning and +22% for storytelling on CIDEr score). Project page: https://mccartney01.github.io/SAM.

Foundation models, pre-trained on massive datasets, have achieved unprecedented generalizability. However, is it truly necessary to involve such vast amounts of data in pre-training, consuming extensive computational resources? This paper introduces data-effective learning, aiming to use data in the most impactful way to pre-train foundation models. This involves strategies that focus on data quality rather than quantity, ensuring the data used for training has high informational value. Data-effective learning plays a profound role in accelerating foundation model training, reducing computational costs, and saving data storage, which is very important as the volume of medical data in recent years has grown beyond many people's expectations. However, due to the lack of standards and comprehensive benchmarks, research on medical data-effective learning is poorly studied. To address this gap, our paper introduces a comprehensive benchmark specifically for evaluating data-effective learning in the medical field. This benchmark includes a dataset with millions of data samples from 31 medical centers (DataDEL), a baseline method for comparison (MedDEL), and a new evaluation metric (NormDEL) to objectively measure data-effective learning performance. Our extensive experimental results show the baseline MedDEL can achieve performance comparable to the original large dataset with only 5% of the data. Establishing such an open data-effective learning benchmark is crucial for the medical foundation model research community because it facilitates efficient data use, promotes collaborative breakthroughs, and fosters the development of cost-effective, scalable, and impactful healthcare solutions.

Large Vision-Language Models (LVLMs) exhibit remarkable capabilities but struggle with ''hallucinations''-inconsistencies between images and their descriptions. Previous hallucination evaluation studies on LVLMs have identified hallucinations in terms of objects, attributes, and relations but overlooked complex hallucinations that create an entire narrative around a fictional entity. In this paper, we introduce a refined taxonomy of hallucinations, featuring a new category: Event Hallucination. We then utilize advanced LLMs to generate and filter fine-grained hallucinatory data consisting of various types of hallucinations, with a particular focus on event hallucinations, laying the groundwork for integrating discriminative and generative evaluation methods within our universal evaluation framework. The proposed benchmark distinctively assesses LVLMs' ability to tackle a broad spectrum of hallucinations, making it a reliable and comprehensive tool for gauging LVLMs' efficacy in handling hallucinations. We will release our code and data.

Streaming videos from resource-constrained front-end devices over networks to resource-rich cloud servers has long been a common practice for surveillance and analytics. Most existing live video analytics (LVA) systems, however, have been built over terrestrial networks, limiting their applications during natural disasters and in remote areas that desperately call for real-time visual data delivery and scene analysis. With the recent advent of space networking, in particular, Low Earth Orbit (LEO) satellite constellations such as Starlink, high-speed truly global Internet access is becoming available and affordable. This paper examines the challenges and potentials of LVA over modern LEO satellite networking (LSN). Using Starlink as the testbed, we have carried out extensive in-the-wild measurements to gain insights into its achievable performance for LVA. The results reveal that the uplink bottleneck in today's LSN, together with the volatile network conditions, can significantly affect the service quality of LVA and necessitate prompt adaptation. We accordingly develop StarStream, a novel LSN-adaptive streaming framework for LVA. At its core, StarStream is empowered by a Transformer-based network performance predictor tailored for LSN and a content-aware configuration optimizer. We discuss a series of key design and implementation issues of StarStream and demonstrate its effectiveness and superiority through trace-driven experiments with real-world network and video processing data.

Recently, there has been a surge in machine-generated natural language content being misused by unauthorized parties. Watermarking is a well-recognized technique to address the issue by tracing the provenance of the text. However, we found that most existing watermarking systems for texts are subject to ad hoc design and thus suffer from fundamental vulnerabilities.

The rapid evolution of egocentric video analysis brings new insights into understanding human activities and intentions from a first-person perspective. Despite this progress, the fragmentation in tasks like action recognition, procedure learning, and moment retrieval, \etc, coupled with inconsistent annotations and isolated model development, hinders a holistic interpretation of video content. In response, we introduce the EAGLE (Egocentric AGgregated Language-video Engine) model and the EAGLE-400K dataset to provide a unified framework that integrates various egocentric video understanding tasks. EAGLE-400K, the \textit{first} large-scale instruction-tuning dataset tailored for egocentric video, features 400K diverse samples to enhance a broad spectrum of tasks from activity recognition to procedure knowledge learning. Moreover, EAGLE, a strong video multimodal large language model (MLLM), is designed to effectively capture both spatial and temporal information. In addition, we propose a set of evaluation metrics designed to facilitate a thorough assessment of MLLM for egocentric video understanding. Our extensive experiments demonstrate EAGLE's superior performance over existing models, highlighting its ability to balance task-specific understanding with holistic video interpretation. With EAGLE, we aim to pave the way for research opportunities and practical applications in real-world scenarios.

Human motion prediction is crucial for human-centric multimedia understanding and interacting. Current methods typically rely on ground truth human poses as observed input, which is not practical for real-world scenarios where only raw visual sensor data is available. To implement these methods in practice, a pre-phrase of pose estimation is essential. However, such two-stage approaches often lead to performance degradation due to the accumulation of errors. Moreover, reducing raw visual data to sparse keypoint representations significantly diminishes the density of information, resulting in the loss of fine-grained features. In this paper, we propose LiDAR-HMP, the first single-LiDAR-based 3D human motion prediction approach, which receives the raw LiDAR point cloud as input and forecasts future 3D human poses directly. Building upon our novel structure-aware body feature descriptor, LiDAR-HMP adaptively maps the observed motion manifold to future poses and effectively models the spatial-temporal correlations of human motions for further refinement of prediction results. Extensive experiments show that our method achieves state-of-the-art performance on two public benchmarks and demonstrates remarkable robustness and efficacy in real-world deployments. https://4dvlab.github.io/project_page/LiDARHMP.html

Current research in food analysis primarily concentrates on tasks such as food recognition, recipe retrieval and nutrition estimation from a single image. Nevertheless, there is a significant gap in exploring the impact of food intake on physiological indicators (e.g., weight) over time. This paper addresses this gap by introducing the DietDiary dataset, which encompasses daily dietary diaries and corresponding weight measurements of real users. Furthermore, we propose a novel task of weight prediction with a dietary diary that aims to leverage historical food intake and weight to predict future weights. To tackle this task, we propose a model-agnostic time series forecasting framework. Specifically, we introduce a Unified Meal Representation Learning (UMRL) module to extract representations for each meal. Additionally, we design a diet-aware loss function to associate food intake with weight variations. By conducting experiments on the DietDiary dataset with two state-of-the-art time series forecasting models, NLinear and iTransformer, we demonstrate that our proposed framework achieves superior performance compared to the original models. We make our dataset, code, and models publicly available at: https://yxg1005.github.io/weight-prediction.

Diffusion-based text-to-image generation models have significantly advanced the field of art content synthesis. However, current portrait stylization methods generally require either model fine-tuning based on examples or the employment of DDIM Inversion to revert images to noise space, both of which substantially decelerate the image generation process. To overcome these limitations, this paper presents an inversion-free portrait stylization framework based on diffusion models that accomplishes content and style feature fusion in merely four sampling steps. We observed that Latent Consistency Models employing consistency distillation can effectively extract representative Consistency Features from noisy images. To blend the Consistency Features extracted from both content and style images, we introduce a Style Enhancement Attention Control technique that meticulously merges content and style features within the attention space of the target image. Moreover, we propose a feature merging strategy to amalgamate redundant features in Consistency Features, thereby reducing the computational load of attention control. Extensive experiments have validated the effectiveness of our proposed framework in enhancing stylization efficiency and fidelity. The code is available at \url{https://github.com/liujin112/ZePo}.

Existing image inpainting methods have achieved remarkable accomplishments in generating visually appealing results, often accompanied by a trend toward creating more intricate structural textures. However, while these models excel at creating more realistic image content, they often leave noticeable traces of tampering, posing a significant threat to security. In this work, we take the anti-forensic capabilities into consideration, firstly proposing an end-to-end training framework for anti-forensic image inpainting named SafePaint. Specifically, we innovatively formulated image inpainting as two major tasks: semantically plausible content completion and region-wise optimization. The former is similar to current inpainting methods that aim to restore the missing regions of corrupted images. The latter, through domain adaptation, endeavors to reconcile the discrepancies between the inpainted region and the unaltered area to achieve anti-forensic goals. Through comprehensive theoretical analysis, we validate the effectiveness of domain adaptation for anti-forensic performance. Furthermore, we meticulously crafted a region-wise separated attention (RWSA) module, which not only aligns with our objective of anti-forensics but also enhances the performance of the model. Extensive qualitative and quantitative evaluations show our approach achieves comparable results to existing image inpainting methods while offering anti-forensic capabilities not available in other methods.

Advanced mobile computing has led to a surge in the need for practical super-resolution (SR) techniques. The look-up table (LUT) based SR-LUT has pioneered a new avenue of research without needing hardware acceleration. Nevertheless, all preceding methods that drew inspiration from the SR-LUT framework invariably resort to interpolation and rotation techniques for diminishing the LUT size, thereby prolonging the inference time and contradicting the original objective of efficient SR. Recently, a study named EC-LUT proposed an expanded convolution method to avoid interpolation operations. However, the performance of EC-LUT regarding SR quality and LUT volume is unsatisfactory. To address these limitations, this paper proposes a novel expanded convolutional neural network (ECNN). Specifically, we further extend feature fusion to the feature channel dimension to enhance mapping ability. In addition, our approach reduces the number of single indexed pixels to just one, eliminating the need for rotation tricks and dramatically reducing the LUT size from the MB level to the KB level, thus improving cache hit rates. By leveraging these improvements, we can stack expanded convolutional layers to form an ECNN, with each layer convertible to LUTs during inference. Experiments show that our method improves the overall performance of the upper limit of LUT based methods. For example, under comparable SR quality conditions, our model achieves state-of-the-art performance in speed and LUT volume.

Existing few-shot learning methods generally focus on designing exquisite structures of meta-learners for learning task-specific prior to improve the discriminative ability of global embeddings. However, they often ignore the importance of learning stability in meta-training, making it difficult to obtain a relatively optimal model. From this key observation, we propose an innovative generic differentiable Reinforcement Learning (RL) strategy for few-shot classification. It aims to explore stable meta-optimization patterns in meta-training by learning generalizable optimizations for producing task-adaptive embeddings. Accordingly, our differentiable RL strategy models the embedding procedure of feature transformation layers in meta-learner to optimize the gradient flow implicitly. Also, we propose a memory module to associate historical and current task states and actions for exploring inter-task similarity. Notably, our RL-based strategy can be easily extended to various backbones. In addition, we propose a novel task state encoder to encode task representation, which fully explores inner-task similarities between support set and query set. Extensive experiments verify that our approach can improve the performance of different backbones and achieve promising results against state-of-the-art methods in few-shot classification.

In this paper, we investigate the challenge of spatio-temporal video prediction, which involves generating future videos based on historical data streams. Existing approaches typically utilize external information such as semantic maps to enhance video prediction, which often neglect the inherent physical knowledge embedded within videos. Furthermore, their high computational demands could impede their applications for high-resolution videos. To address these constraints, we introduce a novel approach called Physics-assisted Spatio-temporal Network (PastNet) for generating high-quality video prediction. The core of our PastNet lies in incorporating a spectral convolution operator in the Fourier domain, which efficiently introduces inductive biases from the underlying physical laws. Additionally, we employ a memory bank with the estimated intrinsic dimensionality to discretize local features during the processing of complex spatio-temporal signals, thereby reducing computational costs and facilitating efficient high-resolution video prediction. Extensive experiments on various widely-used datasets demonstrate the effectiveness and efficiency of the proposed PastNet compared with a range of state-of-the-art methods, particularly in high-resolution scenarios. Our code is available at https://github.com/easylearningscores/PastNet.