ACM MM 2024 Papers

Embodied Everyday Task is a popular task in the embodied AI community, requiring agents to make a sequence of actions based on natural language instructions and visual observations. Traditional learning-based approaches face two challenges. Firstly, natural language instructions often lack explicit task planning. Secondly, extensive training is required to equip models with knowledge of the task environment. Previous works based on Large Language Model (LLM) either suffer from poor performance due to the lack of task-specific knowledge or rely on ground truth as few-shot samples. To address the above limitations, we propose a novel approach called Progressive Retrieval Augmented Generation (P-RAG), which not only effectively leverages the powerful language processing capabilities of LLMs but also progressively accumulates task-specific knowledge without ground-truth. Compared to the conventional RAG methods, which retrieve relevant information from the database in a one-shot manner to assist generation, P-RAG introduces an iterative approach to progressively update the database. In each iteration, P-RAG retrieves the latest database and obtains historical information from the previous interaction as experiential references for the current interaction. Moreover, we also introduce a more granular retrieval scheme that not only retrieves similar tasks but also incorporates retrieval of similar situations to provide more valuable reference experiences. Extensive experiments reveal that P-RAG achieves competitive results without utilizing ground truth and can even further improve performance through self-iterations.

ControlNet excels at creating content that closely matches precise contours in user-provided masks. However, when these masks contain noise, as a frequent occurrence with non-expert users, the output would include unwanted artifacts. This paper first highlights the crucial role of controlling the impact of these inexplicit masks with diverse deterioration levels through in-depth analysis. Subsequently, to enhance controllability with inexplicit masks, an advanced Shape-aware ControlNet consisting of a deterioration estimator and a shape-prior modulation block is devised. The deterioration estimator assesses the deterioration factor of the provided masks. Then this factor is used in a modulation block to adaptively adjust the model's contour-following ability, which helps it dismiss the noise part in the inexplicit masks. Extensive experiments prove its effectiveness in encouraging ControlNet to interpret inaccurate spatial conditions robustly rather than blindly following the given contours, suitable for diverse kinds of conditions. We showcase application scenarios like modifying shape priors and composable shape-controllable generation. Codes are available at github.

While neural radiance fields (NeRF) have shown promise in novel view synthesis, their implicit representation limits explicit control over object manipulation. Existing research has proposed the integration of explicit geometric proxies to enable deformation. However, these methods face two primary challenges: firstly, the time-consuming and computationally demanding tetrahedralization process; and secondly, handling complex or thin structures often leads to either excessive, storage-intensive tetrahedral meshes or poor-quality ones that impair deformation capabilities. To address these challenges, we propose DeformRF, a method that seamlessly integrates the manipulability of tetrahedral meshes with the high-quality rendering capabilities of feature grid representations. To avoid ill-shaped tetrahedra and tetrahedralization for each object, we propose a two-stage training strategy. Starting with an almost-regular tetrahedral grid, our model initially retains key tetrahedra surrounding the object and subsequently refines object details using finer-granularity mesh in the second stage. We also present the concept of recursively subdivided tetrahedra to create higher-resolution meshes implicitly. This enables multi-resolution encoding while only necessitating the storage of the coarse tetrahedral mesh generated in the first training stage. We conduct a comprehensive evaluation of our DeformRF on both synthetic and real-captured datasets. Both quantitative and qualitative results demonstrate the effectiveness of our method for novel view synthesis and deformation tasks. Project page: https://ustc3dv.github.io/DeformRF/

Image matching aims at identifying corresponding points between a pair of images. Currently, detector-free methods have shown impressive performance in challenging scenarios, thanks to their capability of generating dense matches and global receptive field. However, performing feature interaction and proposing matches across the entire image is unnecessary, because not all image regions contribute to the matching process. Interacting and matching in unmatchable areas can introduce errors, reducing matching accuracy and efficiency. Meanwhile, the scale discrepancy issue still troubles existing methods. To address above issues, we propose PRogressive dependency maxImization for Scale-invariant image Matching (PRISM), which jointly prunes irrelevant patch features and tackles the scale discrepancy. To do this, we firstly present a Multi-scale Pruning Module (MPM) to adaptively prune irrelevant features by maximizing the dependency between the two feature sets. Moreover, we design the Scale-Aware Dynamic Pruning Attention (SADPA) to aggregate information from different scales via a hierarchical design. Our method's superior matching performance and generalization capability are confirmed by leading accuracy across various evaluation benchmarks and downstream tasks. The code is publicly available at https://github.com/Master-cai/PRISM.

Traditional dense video captioning predominantly focuses on edited exocentric footage. These videos are filmed from an external perspective and generally feature distinct transitions between different events, as exemplified in edited instructional videos. However, such videos do not genuinely reflect the way we perceive our real lives. Instead, we observe the world from an egocentric viewpoint and witness only continuous unedited footage. To facilitate further research, we introduce a new topic: Egocentric Vehicle Dense Video Captioning, in classic vehicle driving scenarios. This is a multi-modal, multi-task subject endeavor for a comprehensive understanding of untrimmed, egocentric driving videos. It consists of three sub-tasks that concentrate on event localization, captioning, and vehicle state estimation separately. To accomplish these tasks, it is necessary to deal with at least three challenges: extracting ego-motion relevant information, describing driving behavior and analyzing the underlying rationale, as well as resolving the boundary ambiguity problem. In response, we devise corresponding solutions, including a vehicle ego-motion learning strategy and a novel adjacent contrastive learning strategy, which effectively address the aforementioned issues. We validate our method by conducting extensive experiments on the BDD-X dataset, all of which show promising results and achieve new state-of-the-art performance on most metrics, which proves the effect of our approach.

Cross-modal (e.g. image-text, video-text) retrieval is an important task in information retrieval and multimodal vision-language understanding field. Temporal understanding makes video-text retrieval more challenging than image-text retrieval. However, we find that the widely used video-text benchmarks have shortcomings in comprehensively assessing abilities of models, especially in temporal understanding, causing large-scale image-text pre-trained models can already achieve comparable zero-shot performance with video-text pre-trained models. In this paper, we introduce RTime, a novel temporal-emphasized video-text retrieval dataset. We first obtain videos of actions or events with significant temporality, and then reverse these videos to create harder negative samples. We then recruit annotators to judge the significance and reversibility of candidate videos, and write captions for qualified videos. We further adopt GPT-4 to extend more captions based on human-written captions. Our RTime dataset currently consists of 21k videos with 10 captions per video, totalling about 122 hours. Based on RTime, we propose three retrieval benchmark tasks: RTime-Origin, RTime-Hard, and RTime-Binary. We further enhance the use of harder-negatives in model training, and benchmark a variety of video-text models on RTime. Extensive experiment analysis proves that RTime indeed poses new and higher challenges to video-text retrieval. We release our RTime dataset https://github.com/qyr0403/Reversed-in-Time to further advance video-text retrieval and multimodal understanding research.

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.

This paper delves into federated class-incremental learning (FCiL), where new classes appear continually or even privately to local clients. However, existing FCiL methods suffer from the problem of spatial-temporal catastrophic forgetting, i.e., forgetting the previously learned knowledge over time and the client-specific information owned by different clients. Additionally, private class and knowledge heterogeneity amongst local clients further exacerbate spatial-temporal forgetting, making FCiL challenging to apply. To address these issues, we propose Federated Class-specific Binary Classifier (FedCBC), an innovative approach to transferring and fusing knowledge across both temporal and spatial perspectives. FedCBC consists of two novel components: (1) continual personalization that distills previous knowledge from a global model to multiple local models, and (2) selective knowledge fusion that enhances knowledge integration of the same class from divergent clients and shares private knowledge with other clients. Extensive experiments using three newly-formulated metrics (termed GA, KRS, and KRT) demonstrate the effectiveness of the proposed approach. Our code is now hosted at: https://github.com/SkyOfBeginning/FedCBC.

As an important task in multimodal information extraction, Multimodal Named Entity Recognition (MNER) has recently attracted considerable attention. One key challenge of MNER lies in the lack of sufficient fine-grained annotated data, especially in low-resource scenarios. Although data augmentation is a widely used technique to tackle the above issue, it is challenging to simultaneously generate synthetic text-image pairs and their corresponding high-quality entity annotations. In this work, we propose a novel Generative Multimodal Data Augmentation (GMDA) framework for MNER, which contains two stages: Multimodal Text Generation and Multimodal Image Generation. Specifically, we first transform each annotated sentence into a linearized labeled sequence, and then train a Label-aware Multimodal Large Language Model (LMLLM) to generate the labeled sequence based on a label-aware prompt and its associated image. We further employ a Stable Diffusion model to generate the synthetic images that are semantically related to these sentences. Experimental results on three benchmark datasets demonstrate the effectiveness of the proposed GMDA framework, which consistently boosts the performance of several competitive methods for two subtasks of MNER in both full-supervision and low-resource settings. The low-resource dataset and source code are released at https://github.com/NUSTM/GMDA.

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 Stable 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.

Video Question Answering (VideoQA) aims to answer natural language questions based on the information observed in videos. Despite the recent success of Large Multimodal Models (LMMs) in image-language understanding, they deal with VideoQA insufficiently, by simply taking uniformly sampled frames as visual inputs, which ignores question-relevant visual clues. Moreover, there are no human annotations for question-critical timestamps in existing VideoQA datasets. In light of this, we propose a weakly supervised framework to enforce the LMMs to reason out the answers with question-critical moments as visual inputs. Specifically, we fuse the question and answer pairs as event descriptions to find multiple keyframes as target moments and pseudo-labels. With these pseudo-labeled keyframes as additionally weak supervision, we devise a lightweight Gaussian-based Contrastive Grounding (GCG) module. GCG learns multiple Gaussian masks to characterize the temporal structure of the video, and sample question-critical frames as positive moments to be the visual inputs of LMMs. Extensive experiments on several benchmarks verify the effectiveness of our framework, and we achieve substantial improvements compared to previous state-of-the-art methods.

Recent progress in large-scale zero-shot speech synthesis has been significantly advanced by language models and diffusion models. However, the generation process of both methods is slow and computationally intensive. Efficient speech synthesis using a lower computing budget to achieve quality on par with previous work remains a significant challenge. In this paper, we present FlashSpeech, a large-scale zero-shot speech synthesis system with approximately 5% of the inference time compared with previous work. FlashSpeech is built on the latent consistency model and applies a novel adversarial consistency training approach that can train from scratch without the need for a pre-trained diffusion model as the teacher. Furthermore, a new prosody generator module enhances the diversity of prosody, making the rhythm of the speech sound more natural. The generation processes of FlashSpeech can be achieved efficiently with one or two sampling steps while maintaining high audio quality and high similarity to the audio prompt for zero-shot speech generation. Our experimental results demonstrate the superior performance of FlashSpeech. Notably, FlashSpeech can be about 20 times faster than other zero-shot speech synthesis systems while maintaining comparable performance in terms of voice quality and similarity. Furthermore, FlashSpeech demonstrates its versatility by efficiently performing tasks like voice conversion, speech editing, and diverse speech sampling. Audio samples can be found in https://flashspeech.github.io/

Multi-view clustering has garnered attention for its effectiveness in addressing heterogeneous data by unsupervisedly revealing underlying correlations between different views. As a mainstream method, multi-view graph clustering has attracted increasing attention in recent years. Despite its success, it still has some limitations. Notably, many methods construct the similarity graph without considering the local geometric structure and exploit coarse-grained complementary and consensus information from different views at the view level. To solve the shortcomings, we focus on local structure consistency and fine-grained representations across multiple views. Specifically, each view's local consistency similarity graph is obtained through the adaptive neighbor. Subsequently, the multi-view similarity tensor is rotated and sliced into fine-grained instance-wise slices. Finally, these slices are fused into the final similarity matrix. Consequently, cross-view consistency can be captured by exploring the intersections of multiple views in an instance-wise manner. We design a collaborative framework with the augmented Lagrangian method to refine all subtasks towards optimal solutions iteratively. Extensive experiments on several multi-view datasets confirm the significant enhancement in clustering accuracy achieved by our method.

Due to the small size of valid samples, multi-source EEG features with high dimensionality can easily cause problems such as overfitting and poor real-time performance of the emotion recognition classifier. Feature selection has been demonstrated as an effective means to solve these problems. Current EEG feature selection research assumes that all dimensions of emotional labels are complete. However, owing to the open acquisition environment, subjective variability, and border ambiguity of individual perceptions of emotion, the training data in the practical application often includes missing information, i.e., multi-dimensional emotional labels of several instances are incomplete. The aforementioned incomplete information directly restricts the accurate construction of the EEG feature selection model for multi-dimensional emotion recognition. To wrestle with the aforementioned problem, we propose a novel EEG feature selection model with weighted self-expression learning (WSEL). The model utilizes self-representation learning and least squares regression to reconstruct the label space through the second-order correlation and higher-order correlation within the multi-dimensional emotional labels and simultaneously realize the EEG feature subset selection under the incomplete information. We have utilized two multimedia-induced emotion datasets with EEG recordings, DREAMER and DEAP, to confirm the effectiveness of WSEL in the missing multi-dimensional emotional feature selection challenge. Compared to nine state-of-the-art feature selection approaches, the experimental results demonstrate that the EEG feature subsets chosen by WSEL can achieve optimal performance in terms of six performance metrics.

Emotion cause analysis has attracted increasing attention in recent years. However, the integration of multimodal information with emotion causes remains underexplored. Existing studies merely extract utterances from conversations as cause evidence, which is too coarse-grained to locate the exact causes from other modalities, especially those that may be reflected only in a specific video frame of an utterance. To address these limitations, we introduce a new task named Multimodal Emotion Cause Generation in Conversations (MECGC), which aims to generate an abstractive summary clearly and intuitively describing the causes that trigger the given emotion based on the multimodal context of conversations. We accordingly construct a dataset named ECGF that contains 1,374 conversations and 7,690 emotion instances from TV series. We further develop a generative framework that first generates emotion-cause aware video captions (Observe) and then facilitates the generation of emotion causes (Generate). The captioning model is trained with examples synthesized by a Multimodal Large Language Model (MLLM). Experimental results demonstrate the effectiveness of our framework and the significance of multimodal information for emotion cause analysis. Our dataset and source codes are available at https://github.com/NUSTM/MECGC.

Speech-preserving Facial Expression Manipulation (SPFEM) aims to alter facial emotions in video content while preserving the facial movements associated with speech. Current works often fall short due to the inadequate representation of emotion as well as the absence of time-aligned paired data-two corresponding frames from the same speaker that showcase the same speech content but differ in emotional expression. In this work, we introduce a novel framework, Self-Supervised Emotion Representation Disentanglement (SSERD), to disentangle emotion representation for accurate emotion transfer while implementing a paired data construction module to facilitate automated, photorealistic facial animations. Specifically, We developed a module for learning emotion latent codes using StyleGAN's latent space, employing a cross-attention mechanism to extract and predict emotion editing codes, with contrastive learning to differentiate emotions. To overcome the lack of strictly paired data in the SPFEM task, we exploit pretrained StyleGAN to generate paired data, focusing on expression vectors unrelated to mouth shape. Additionally, we employed a hybrid training strategy using both synthetic paired and real unpaired data to enhance the realism of SPFEM model's generated images. Extensive experiments conducted on benchmark datasets, including MEAD and RAVDESS, have validated the effectiveness of our framework, demonstrating its superior capability in generating photorealistic and expressive facial animations.

Recently, indoor 3D object detection has shown impressive progress. However, these improvements have come at the cost of increased memory consumption and longer inference times, making it difficult to apply these methods in practical scenarios. To address this issue, knowledge distillation has emerged as a promising technique for model acceleration. In this paper, we propose the VRDistill framework, the first knowledge distillation framework designed for efficient indoor 3D object detection. Our VRDistill framework includes a refinement module and a soft foreground mask operation to enhance the quality of the distillation. The refinement module utilizes trainable layers to improve the quality of the teacher's votes, while the soft foreground mask operation focuses on foreground votes, further enhancing the distillation performance. Comprehensive experiments on the ScanNet and SUN-RGBD datasets demonstrate the effectiveness and generalization ability of our VRDistill framework.