Continual pre-training (CPT) has been an important approach for adapting language models to specific domains or tasks. In this paper, we comprehensively study its key designs to balance the new abilities while retaining the original abilities, and present an effective CPT method that can greatly improve the Chinese language ability and scientific reasoning ability of LLMs. To achieve it, we design specific data mixture and curriculum strategies based on existing datasets and synthetic high-quality data. Concretely, we synthesize multidisciplinary scientific QA pairs based on related web pages to guarantee the data quality, and also devise the performance tracking and data mixture adjustment strategy to ensure the training stability. For the detailed designs, we conduct preliminary studies on a relatively small model, and summarize the findings to help optimize our CPT method. Extensive experiments on a number of evaluation benchmarks show that our approach can largely improve the performance of Llama-3 (8B), including both the general abilities (+8.81 on C-Eval and +6.31 on CMMLU) and the scientific reasoning abilities (+12.00 on MATH and +4.13 on SciEval). Our model, data, and codes are available at https://github.com/RUC-GSAI/Llama-3-SynE.

In recent years, numerous tasks have been proposed to encourage model to develop specified capability in understanding audio-visual scene, primarily categorized into temporal localization, spatial localization, spatio-temporal reasoning, and pixel-level understanding. Instead, human possesses a unified understanding ability for diversified tasks. Therefore, designing an audio-visual model with general capability to unify these tasks is of great value. However, simply joint training for all tasks can lead to interference due to the heterogeneity of audiovisual data and complex relationship among tasks. We argue that this problem can be solved through explicit cooperation among tasks. To achieve this goal, we propose a unified learning method which achieves explicit inter-task cooperation from both the perspectives of data and model thoroughly. Specifically, considering the labels of existing datasets are simple words, we carefully refine these datasets and construct an Audio-Visual Unified Instruction-tuning dataset with Explicit reasoning process (AV-UIE), which clarifies the cooperative relationship among tasks. Subsequently, to facilitate concrete cooperation in learning stage, an interaction-aware LoRA structure with multiple LoRA heads is designed to learn different aspects of audiovisual data interaction. By unifying the explicit cooperation across the data and model aspect, our method not only surpasses existing unified audio-visual model on multiple tasks, but also outperforms most specialized models for certain tasks. Furthermore, we also visualize the process of explicit cooperation and surprisingly find that each LoRA head has certain audio-visual understanding ability. Code and dataset: https://github.com/GeWu-Lab/Crab

Sensory training during the early ages is vital for human development. Inspired by this cognitive phenomenon, we observe that the early training stage is also important for the multimodal learning process, where dataset information is rapidly acquired. We refer to this stage as the prime learning window. However, based on our observation, this prime learning window in multimodal learning is often dominated by information-sufficient modalities, which in turn suppresses the information acquisition of informationinsufficient modalities. To address this issue, we propose Information Acquisition Regulation (InfoReg), a method designed to balance information acquisition among modalities. Specifically, InfoReg slows down the information acquisition process of information-sufficient modalities during the prime learning window, which could promote information acquisition of information-insufficient modalities. This regulation enables a more balanced learning process and improves the overall performance of the multimodal network. Experiments show that InfoReg outperforms related multimodal imbalanced methods across various datasets, achieving superior model performance. The code is available at https://github.com/GeWu-Lab/InfoReg_CVPR2025.

High-quality image captions play a crucial role in improving the performance of cross-modal applications such as text-to-image generation, text-to-video generation, and text-image retrieval. To generate long-form, high-quality captions, many recent studies have employed multimodal large language models (MLLMs). However, current MLLMs often produce captions that lack fine-grained details or suffer from hallucinations, a challenge that persists in both open-source and closed-source models. Inspired by Feature-Integration theory, which suggests that attention must focus on specific regions to integrate visual information effectively, we propose a divide-then-aggregate strategy. Our method first divides the image into semantic and spatial patches to extract fine-grained details, enhancing the model’s local perception of the image. These local details are then hierarchically aggregated to generate a comprehensive global description. To address hallucinations and inconsistencies in the generated captions, we apply a semantic-level filtering process during hierarchical aggregation. This training-free pipeline can be applied to both open-source models (LLaVA-1.5, LLaVA-1.6, Mini-Gemini) and closed-source models (Claude-3.5-Sonnet, GPT-4o, GLM-4V-Plus). Extensive experiments demonstrate that our method generates more detailed, reliable captions, advancing multimodal description generation without requiring model retraining. The source code are available at https://github.com/GeWu-Lab/Patch-Matters

Building a generalizable self-correction system is crucial for robots to recover from failures. Despite advancements in Multimodal Large Language Models (MLLMs) that empower robots with semantic reflection ability for failure, translating semantic reflection into "how to correct" fine-grained robotic actions remains a significant challenge. To address this gap, we build the Phoenix framework, which leverages motion instruction as a bridge to connect high-level semantic reflection with low-level robotic action correction. In this motion-based self-reflection framework, we start with a dual-process motion adjustment mechanism with MLLMs to translate the semantic reflection into coarse-grained motion instruction adjustment. To leverage this motion instruction for guiding "how to correct" fine-grained robotic actions, a multi-task motion-conditioned diffusion policy is proposed to integrate visual observations for high-frequency robotic action correction. By combining these two models, we could shift the demand for generalization capability from the low-level manipulation policy to the MLLMs-Driven motion adjustment model and facilitate precise, fine-grained robotic action correction. Utilizing this framework, we further develop a lifelong learning method to automatically improve the model’s capability from interactions with dynamic environments. The experiments conducted in both the RoboMimic simulation and real-world scenarios prove the superior generalization and robustness of our framework across a variety of manipulation tasks. Our code is released at https://github.com/GeWu-Lab/Motion-Based-Self-Reflection-Framework.

Visuo-tactile sensors aim to emulate human tactile perception, enabling robots to precisely understand and manipulate objects. Over time, numerous meticulously designed visuo-tactile sensors have been integrated into robotic systems, aiding in completing various tasks. However, the distinct data characteristics of these low-standardized visuo-tactile sensors hinder the establishment of a powerful tactile perception system. We consider that the key to addressing this issue lies in learning unified multi-sensor representations, thereby integrating the sensors and promoting tactile knowledge transfer between them. To achieve unified representation of this nature, we introduce TacQuad, an aligned multi-modal multi-sensor tactile dataset from four different visuo-tactile sensors, which enables the explicit integration of various sensors. Recognizing that humans perceive the physical environment by acquiring diverse tactile information such as texture and pressure changes, we further propose to learn unified multi-sensor representations from both static and dynamic perspectives. By integrating tactile images and videos, we present AnyTouch, a unified static-dynamic multi-sensor representation learning framework with a multi-level structure, aimed at both enhancing comprehensive perceptual abilities and enabling effective cross-sensor transfer. This multi-level architecture captures pixel-level details from tactile data via masked modeling and enhances perception and transferability by learning semantic-level sensor-agnostic features through multi-modal alignment and cross-sensor matching. We provide a comprehensive analysis of multi-sensor transferability, and validate our method on various offline datasets and in the real-world pouring task. Experimental results show that our method outperforms existing methods, exhibits outstanding static and dynamic perception capabilities across various sensors. The code, TacQuad dataset and AnyTouch model are fully available at gewu-lab.github.io/AnyTouch/.

Multimodal learning faces challenges in effectively fusing information from diverse modalities, especially when modality quality varies across samples. Dynamic fusion strategies, such as attention mechanism in Transformers, aim to address such challenge by adaptively emphasizing modalities based on the characteristics of input data. However, through amounts of carefully designed experiments, we surprisingly observed that the dynamic adaptability of widely-used self-attention models diminishes. Model tends to prefer one modality regardless of data characteristics. This bias triggers a self-reinforcing cycle that progressively overemphasizes the favored modality, widening the distribution gap in attention keys across modalities and deactivating attention mechanism’s dynamic properties. To revive adaptability, we propose a simple yet effective method Rolling Query (RollingQ), which balances attention allocation by rotating the query to break the self-reinforcing cycle and mitigate the key distribution gap. Extensive experiments on various multimodal scenarios validate the effectiveness of RollingQ and the restoration of cooperation dynamics is pivotal for enhancing the broader capabilities of widely deployed multimodal Transformers. The source code is available at https://github.com/GeWu-Lab/RollingQ_ICML2025.

Interactions between modalities—redundancy, uniqueness, and synergy—collectively determine the composition of multimodal information. Understanding these interactions is crucial for analyzing information dynamics in multimodal systems, yet their accurate sample-level quantification presents significant theoretical and computational challenges. To address this, we introduce the Lightweight Sample-wise Multimodal Interaction (LSMI) estimator, rigorously grounded in pointwise information theory. We first develop a redundancy estimation framework, employing an appropriate pointwise information measure to quantify this most decomposable and measurable interaction. Building upon this, we propose a general interaction estimation method that employs efficient entropy estimation, specifically tailored for sample-wise estimation in continuous distributions. Extensive experiments on synthetic and real-world datasets validate LSMI’s precision and efficiency. Crucially, our sample-wise approach reveals fine-grained sample- and category-level dynamics within multimodal data, enabling practical applications such as redundancy-informed sample partitioning, targeted knowledge distillation, and interaction-aware model ensembling. The code is available at https://github.com/GeWu-Lab/LSMI_Estimator.

The rapid development of short-form video platforms has provided new avenues for advertisement recommendation, where advertisements are typically displayed as a thumbnail button at the beginning of a video and expand after a certain period. The audio-visual content of a video usually impacts the emotions and attention of users, which further partially determines user behaviors after the advertisement expands. This impact varies depending on the moment of the expansion. Hence, the task of predicting the expansion moment for each video is essential to improve the quality of advertisement recommendation, which is the focus of this paper. The main challenge of this task lies in the requirement to understand both audio-visual video content and user behaviors. To facilitate our research on this task, we first collect a large-scale dataset containing 18,901 videos with over 1.58 billion real user-watching records as the annotations. We also propose a quantitative metric for the task. To address the challenge of the task as well as predict a proper expansion moment for each video, we develop a prediction framework. Firstly, we utilize pretrained models to extract features, then refine extracted features via intra-modal and inter-modal interactions to enhance the understanding of the audio-visual content. Secondly, we propose an uncertainty-based logit adjustment strategy, which leverages prior preferences of user behaviors to adjust the output logit of the model adaptively, achieving a better prediction of user behaviors. Extensive experiments demonstrate the effectiveness of our proposed framework. We hope this work could be a meaningful step towards better video understanding and advertisement recommendation in short-form video platforms. Code and dataset are available at https://github.com/GeWu-Lab/WCAE.

In this paper, we reveal that most current efficient multimodal fine-tuning methods are hindered by a key limitation: they are directly borrowed from LLMs, often neglecting the intrinsic differences of multimodal scenarios and even affecting the full utilization of all modalities. Inspired by our empirical observation, we argue that unimodal adaptation and cross-modal adaptation are two essential parts for the effective fine-tuning of MLLMs. From this perspective, we propose Multimodal Low-rank Adaptation (MokA), a multimodal-aware efficient fine-tuning strategy that takes multimodal characteristics into consideration. It compresses unimodal information by modality-specific parameters while explicitly enhancing cross-modal interaction, ensuring both unimodal and cross-modal adaptation. Extensive experiments cover three representative multimodal scenarios (audio-visual-text, visual-text, and speech-text), and multiple LLM backbones (LLaMA2, Qwen2, Qwen2.5-VL, etc). Consistent improvements indicate the efficacy and versatility of the proposed method. Ablation studies and efficiency evaluation are also conducted to fully asses our method. Overall, we think MokA provides a more targeted solution for efficient adaptation of MLLMs, paving the way for further exploration.

Personalized recommender systems, which focus on predicting users' interests, have significantly enhanced user experiences across diverse applications. However, existing approaches implicitly model users' preferences through fitting the fine-grained labels (e.g., click labels), but often neglecting the coarse-grained interest information inherent in the inputs themselves. Relying solely on the fine-grained labels could bring negative impact on interest modeling and limit the performance, as the labels may carry inevitable noise in real-world scenarios. In addition, it is considerably demanding in terms of data for most existing approaches to effectively model users' multi-granularity interests with limited or no supporting examples, resulting in subpar performance due to the significant long-tail phenomenon. To tackle these issues, we propose a novel learning framework named the Multi-Granularity Interest Prediction Framework (MGIPF), for better modeling users' diverse interests. Unlike prior work, our key idea is to utilize both the coarse-grained and fine-grained interests for supervising the training of models. Specifically, we introduce a pseudo-labeling approach explicitly mining users' potential multi-granularity interests from the raw data, and propose coarse-grained interest prediction modules that collaborate to utilize the multi-granularity supervision signals to enhance the learning of low-frequency items. The corresponding coarse-grained losses are softly weighted, taking into account the varying confidence of potential multi-granularity preferences on positive and negative samples. Importantly, our framework is lightweight and adaptable, capable of being applied effectively to mainstream recommendation models, establishing a comprehensive end-to-end training process. Extensive experiments conducted on three publicly available datasets have demonstrated the efficacy of our approach. The code is available at https://github.com/GeWu-Lab/MGIPF.

Multimodal learning is expected to boost model performance by integrating information from different modalities. However, its potential is not fully exploited because the widely-used joint training strategy, which has a uniform objective for all modalities, leads to imbalanced and under-optimized uni-modal representations. Specifically, we point out that there often exists modality with more discriminative information, e.g., vision of playing football and sound of blowing wind. They could dominate the joint training process, resulting in other modalities being significantly under-optimized. To alleviate this problem, we first analyze the under-optimized phenomenon from both the feed-forward and the back-propagation stages during optimization. Then, On-the-fly Prediction Modulation (OPM) and On-the-fly Gradient Modulation (OGM) strategies are proposed to modulate the optimization of each modality, by monitoring the discriminative discrepancy between modalities during training. Concretely, OPM weakens the influence of the dominant modality by dropping its feature with dynamical probability in the feed-forward stage, while OGM mitigates its gradient in the back-propagation stage. In experiments, our methods demonstrate considerable improvement across a variety of multimodal tasks. These simple yet effective strategies not only enhance performance in vanilla and task-oriented multimodal models, but also in more complex multimodal tasks, showcasing their effectiveness and flexibility.

Never having seen an object and heard its sound simultaneously, can the model still accurately localize its visual position from the input audio? In this work, we concentrate on the Audio-Visual Localization and Segmentation tasks but under the demanding zero-shot and few-shot scenarios. To achieve this goal, different from existing approaches that mostly employ the encoder-fusion-decoder paradigm to decode localization information from the fused audio-visual feature, we introduce the encoder-prompt-decoder paradigm, aiming to better fit the data scarcity and varying data distribution dilemmas with the help of abundant knowledge from pre-trained models. Specifically, we first propose to construct a Semantic-aware Audio Prompt (SAP) to help the visual foundation model focus on sounding objects, meanwhile, the semantic gap between the visual and audio modalities is also encouraged to shrink. Then, we develop a Correlation Adapter (ColA) to keep minimal training efforts as well as maintain adequate knowledge of the visual foundation model. By equipping with these means, extensive experiments demonstrate that this new paradigm outperforms other fusion-based methods in both the unseen class and cross-dataset settings. We hope that our work can further promote the generalization study of Audio-Visual Localization and Segmentation in practical application scenarios. Project page: https://github.com/GeWu-Lab/Generalizable-Audio-Visual-Segmentation

One primary topic of multimodal learning is to jointly incorporate heterogeneous information from different modalities. However, most models often suffer from unsatisfactory multimodal cooperation, which cannot jointly utilize all modalities well. Some methods are proposed to identify and enhance the worse learnt modality, but they are often hard to provide the fine-grained observation of multimodal cooperation at sample-level with theoretical support. Hence, it is essential to reasonably observe and improve the fine-grained cooperation between modalities, especially when facing realistic scenarios where the modality discrepancy could vary across different samples. To this end, we introduce a sample-level modality valuation metric to evaluate the contribution of each modality for each sample. Via modality valuation, we observe that modality discrepancy indeed could be different at sample-level, beyond the global contribution discrepancy at dataset-level. We further analyze this issue and improve cooperation between modalities at sample-level by enhancing the discriminative ability of low-contributing modalities in a targeted manner. Overall, our methods reasonably observe the fine-grained uni-modal contribution and achieve considerable improvement. The source code and dataset are available at https://github.com/GeWu-Lab/Valuate-and-Enhance-Multimodal-Cooperation.

Multi-modal models have shown a promising capability to effectively integrate information from various sources, yet meanwhile, they are found vulnerable to pervasive perturbations, such as uni-modal attacks and missing conditions. To counter these perturbations, robust multi-modal representations are highly expected, which are positioned well away from the discriminative multi-modal decision boundary. In this paper, different from conventional empirical studies, we focus on a commonly used joint multi-modal framework and theoretically discover that larger uni-modal representation margins and more reliable integration for modalities are essential components for achieving higher robustness. This discovery can further explain the limitation of multi-modal robustness and the phenomenon that multi-modal models are often vulnerable to attacks on the specific modality. Moreover, our analysis reveals how the widespread issue, that the model has different preferences for modalities, limits the multi-modal robustness by influencing the essential components and could lead to attacks on the specific modality highly effective. Inspired by our theoretical finding, we introduce a training procedure called Certifiable Robust Multi-modal Training (CRMT), which can alleviate this influence from modality preference and explicitly regulate essential components to significantly improve robustness in a certifiable manner. Our method demonstrates substantial improvements in performance and robustness compared with existing methods. Furthermore, our training procedure can be easily extended to enhance other robust training strategies, highlighting its credibility and flexibility.

Multimodal learning methods with targeted unimodal learning objectives have exhibited their superior efficacy in alleviating the imbalanced multimodal learning problem. However, in this paper, we identify the previously ignored gradient conflict between multimodal and unimodal learning objectives, potentially misleading the unimodal encoder optimization. To well diminish these conflicts, we observe the discrepancy between multimodal loss and unimodal loss, where both gradient magnitude and covariance of the easier-to-learn multimodal loss are smaller than the unimodal one. With this property, we analyze Pareto integration under our multimodal scenario and propose MMPareto algorithm, which could ensure a final gradient with direction that is common to all learning objectives and enhanced magnitude to improve generalization, providing innocent unimodal assistance. Finally, experiments across multiple types of modalities and frameworks with dense cross-modal interaction indicate our superior and extendable method performance. Our method is also expected to facilitate multi-task cases with a clear discrepancy in task difficulty, demonstrating its ideal scalability. The source code and dataset are available at https://github.com/GeWu-Lab/MMPareto_ICML2024.

Generalizable articulated object manipulation is essential for home-assistant robots. Recent efforts focus on imitation learning from demonstrations or reinforcement learning in simulation, however, due to the prohibitive costs of real-world data collection and precise object simulation, it still remains challenging for these works to achieve broad adaptability across diverse articulated objects. Recently, many works have tried to utilize the strong in-context learning ability of Large Language Models (LLMs) to achieve generalizable robotic manipulation, but most of these researches focus on high-level task planning, sidelining low-level robotic control. In this work, building on the idea that the kinematic structure of the object determines how we can manipulate it, we propose a kinematic-aware prompting framework that prompts LLMs with kinematic knowledge of objects to generate low-level motion trajectory waypoints, supporting various object manipulation. To effectively prompt LLMs with the kinematic structure of different objects, we design a unified kinematic knowledge parser, which represents various articulated objects as a unified textual description containing kinematic joints and contact location. Building upon this unified description, a kinematic-aware planner model is proposed to generate precise 3D manipulation waypoints via a designed kinematic-aware chain-of-thoughts prompting method. Our evaluation spanned 48 instances across 16 distinct categories, revealing that our framework not only outperforms traditional methods on 8 seen categories but also shows a powerful zero-shot capability for 8 unseen articulated object categories with only 17 demonstrations. Moreover, the real-world experiments on 7 different object categories prove our framework’s adaptability in practical scenarios. Code is released at https://github.com/GeWu-Lab/LLM_articulated_object_manipulation.

3D perception ability is crucial for generalizable robotic manipulation. While recent foundation models have made significant strides in perception and decision-making with RGB-based input, their lack of 3D perception limits their effectiveness in fine-grained robotic manipulation tasks. To address these limitations, we propose a Depth Information Injection (DI2) framework that leverages the RGB-Depth modality for policy fine-tuning, while relying solely on RGB images for robust and efficient deployment. Concretely, we introduce the Depth Completion Module (DCM) to extract the spatial prior knowledge related to depth information and generate virtual depth information from RGB inputs to aid policy deployment. Further, we propose the Depth-Aware Codebook (DAC) to eliminate noise and reduce the cumulative error from the depth prediction. In the inference phase, this framework employs RGB inputs and accurately predicted depth data to generate the manipulation action. We conduct experiments on simulated LIBERO environments and real-world scenarios, and the experiment results prove that our method could effectively enhance the pre-trained RGB-based policy with 3D perception ability for robotic manipulation. The website is released at https://gewu-lab.github.io/DepthHelps-IROS2024.

Community researchers have developed a range of advanced audio-visual segmentation models aimed at improving the quality of sounding objects' masks. While masks created by these models may initially appear plausible, they occasionally exhibit anomalies with incorrect grounding logic. We attribute this to real-world inherent preferences and distributions as a simpler signal for learning than the complex audio-visual grounding, which leads to the disregard of important modality information. Generally, the anomalous phenomena are often complex and cannot be directly observed systematically. In this study, we made a pioneering effort with the proper synthetic data to categorize and analyze phenomena as two types "audio priming bias" and "visual prior" according to the source of anomalies. For audio priming bias, to enhance audio sensitivity to different intensities and semantics, a perception module specifically for audio perceives the latent semantic information and incorporates information into a limited set of queries, namely active queries. Moreover, the interaction mechanism related to such active queries in the transformer decoder is customized to adapt to the need for interaction regulating among audio semantics. For visual prior, multiple contrastive training strategies are explored to optimize the model by incorporating a biased branch, without even changing the structure of the model. During experiments, observation demonstrates the presence and the impact that has been produced by the biases of the existing model. Finally, through experimental evaluation of AVS benchmarks, we demonstrate the effectiveness of our methods in handling both types of biases, achieving competitive performance across all three subsets.

Training multimodal models requires a large amount of labeled data. Active learning (AL) aim to reduce labeling costs. Most AL methods employ warm-start approaches, which rely on sufficient labeled data to train a well-calibrated model that can assess the uncertainty and diversity of unlabeled data. However, when assembling a dataset, labeled data are often scarce initially, leading to a cold-start problem. Additionally, most AL methods seldom address multimodal data, highlighting a research gap in this field. Our research addresses these issues by developing a two-stage method for Multi-Modal Cold-Start Active Learning (MMCSAL). Firstly, we observe the modality gap, a significant distance between the centroids of representations from different modalities, when only using cross-modal pairing information as self-supervision signals. This modality gap affects data selection process, as we calculate both uni-modal and cross-modal distances. To address this, we introduce uni-modal prototypes to bridge the modality gap. Secondly, conventional AL methods often falter in multimodal scenarios where alignment between modalities is overlooked. Therefore, we propose enhancing cross-modal alignment through regularization, thereby improving the quality of selected multimodal data pairs in AL. Finally, our experiments demonstrate MMCSAL's efficacy in selecting multimodal data pairs across three multimodal datasets.

Abstract Fine-tuning deep neural networks pre-trained on large scale datasets is one of the most practical transfer learning paradigm given limited quantity of training samples. To obtain better generalization, using the starting point as the reference (SPAR), either through weights or features, has been successfully applied to transfer learning as a regularizer. However, due to the domain discrepancy between the source and target task, there exists obvious risk of negative transfer in a straightforward manner of knowledge preserving. In this paper, we propose a novel transfer learning algorithm, introducing the idea of Target-awareness REpresentation Disentanglement ( $$\textrm{TRED}$$ TRED ), where the relevant knowledge with respect to the target task is disentangled from the original source model and used as a regularizer during fine-tuning the target model. Two alternative approaches, maximizing Maximum Mean Discrepancy (Max-MMD) and minimizing mutual information (Min-MI) are introduced to achieve the desired disentanglement. Experiments on various real world datasets show that our method stably improves the standard fine-tuning by more than 2% in average. $$\textrm{TRED}$$ TRED also outperforms related state-of-the-art transfer learning regularizers such as $$\mathrm {L^2\text {-}SP}$$ L 2 - SP , $$\textrm{AT}$$ AT , $$\textrm{DELTA}$$ DELTA , and $$\textrm{BSS}$$ BSS . Moreover, our solution is compatible with different choices of disentangling strategies. While the combination of Max-MMD and Min-MI typically achieves higher accuracy, only using Max-MMD can be a preferred choice in applications with low resource budgets.

Audio-visual learning helps to comprehensively under-stand the world by fusing practical information from multiple modalities. However, recent studies show that the imbalanced optimization of uni-modal encoders in a joint-learning model is a bottleneck to enhancing the model’s performance. We further find that the up-to-date imbalance-mitigating methods fail on some audio-visual fine-grained tasks, which have a higher demand for distinguishable feature distribution. Fueled by the success of cosine loss that builds hyperspherical feature spaces and achieves lower intra-class angular variability, this paper proposes Multi-Modal Cosine loss, MMCosine. It performs a modality-wise L2normalization to features and weights towards balanced and better multi-modal fine-grained learning. We demonstrate that our method can alleviate the imbalanced optimization from the perspective of weight norm and fully exploit the discriminability of the cosine metric. Extensive experiments prove the effectiveness of our method and the versatility with advanced multi-modal fusion strategies and up-to-date imbalance-mitigating methods. The project page is https://gewu-lab.github.io/MMCosine/.

Transfer learning has been a popular learning paradigm in the deep learning era, especially in annotation-insufficient scenarios. Better ImageNet pre-trained models have been demonstrated, from the perspective of architecture, by previous research to have better transferability to downstream tasks[26]. However, in this paper, we find that during the same pre-training process, models at middle epochs, which are inadequately pre-trained, can outperform fully trained models when used as feature extractors (FE), while the fine-tuning (FT) performance still grows with the source performance. This reveals that there is not a solid positive correlation between top-1 accuracy on ImageNet and the transferring result on target data. Based on the contradictory phenomenon between FE and FT that a better feature extractor fails to be fine-tuned better accordingly, we conduct comprehensive analyses on features before the softmax layer to provide insightful explanations. Our discoveries suggest that, during pre-training, models tend to first learn spectral components corresponding to large singular values and the residual components contribute more when fine-tuning.

Audio-Visual Question Answering (AVQA) task aims to answer questions about different visual objects, sounds, and their associations in videos. Such naturally multi-modal videos are composed of rich and complex dynamic audio-visual components, where most of which could be unrelated to the given questions, or even play as interference in answering the content of interest. Oppositely, only focusing on the question-aware audio-visual content could get rid of influence, meanwhile enabling the model to answer more efficiently. In this paper, we propose a Progressive Spatio-Temporal Perception Network (PSTP-Net), which contains three modules that progressively identify key spatio-temporal regions w.r.t. questions. Specifically, a temporal segment selection module is first introduced to select the most relevant audio-visual segments related to the given question. Then, a spatial region selection module is utilized to choose the most relevant regions associated with the question from the selected temporal segments. To further refine the selection of features, anaudio-guided visual attention module is employed to perceive the association between auido and selected spatial regions. Finally, the spatio-temporal features from these modules are integrated for answering the question. Extensive experimental results on the public MUSIC-AVQA and AVQA datasets provide compelling evidence of the effectiveness and efficiency of PSTP-Net.

To facilitate the research on intelligent and human-like chatbots with multi-modal context, we introduce a new video-based multi-modal dialogue dataset, called TikTalk. We collect 38K videos from a popular video-sharing platform, along with 367K conversations posted by users beneath them. Users engage in spontaneous conversations based on their multi-modal experiences from watching videos, which helps recreate real-world chitchat context. Compared to previous multi-modal dialogue datasets, the richer context types in TikTalk lead to more diverse conversations, but also increase the difficulty in capturing human interests from intricate multi-modal information to generate personalized responses. Moreover, external knowledge is more frequently evoked in our dataset. These facts reveal new challenges for multi-modal dialogue models. We quantitatively demonstrate the characteristics of TikTalk, propose a video-based multi-modal chitchat task, and evaluate several dialogue baselines. Experimental results indicate that the models incorporating large language models (LLM) can generate more diverse responses, while the model utilizing knowledge graphs to introduce external knowledge performs the best overall. Furthermore, no existing model can solve all the above challenges well. There is still a large room for future improvements, even for LLM with visual extensions. Our dataset is available at https://ruc-aimind.github.io/projects/TikTalk/.

Self-supervised sound source localization in unconstrained visual scenes is an important task of audio-visual learning. In this paper, we propose a visual reasoning module to explicitly exploit the rich visual context semantics, which alleviates the issue of insufficient utilization of visual information in previous works. The learning objectives are carefully designed to provide stronger supervision signals for the extracted visual semantics while enhancing the audio-visual interactions, which lead to more robust feature representations. Extensive experimental results demonstrate that our approach significantly boosts the localization performances on various datasets, even without initializations pretrained on ImageNet. Moreover, with the visual context exploitation, our framework can accomplish both the audio-visual and purely visual inference, which expands the application scope of the sound source localization task and further raises the competitiveness of our approach.

Recent years have witnessed the success of deep learning on the visual sound separation task. However, existing works follow similar settings where the training and testing datasets share the same musical instrument categories, which to some extent limits the versatility of this task. In this work, we focus on a more general and challenging scenario, namely the separation of unknown musical instruments, where the categories in training and testing phases have no direct overlap with each other. To tackle this new setting, we propose the "Separation-with-Consistency" (SeCo) framework, which can accomplish the separation on unknown categories by exploiting the consistency constraints. Furthermore, to capture richer characteristics of the novel melodies, we devise an online matching strategy, which can bring stable enhancements with no cost of extra parameters. Experiments demonstrate that our SeCo framework exhibits strong adaptation ability on the novel musical categories and outperforms the baseline methods by a notable margin.

Due to the difficulty of annotating large amounts of training data, directly learning the association of sound and its makers in natural videos is a challenging task for machines. In this paper, we present a novel audiovisual model that introduces a soft-clustering module as the audio and visual content detector, and regards the pervasive property of audiovisual concurrency as the latent supervision for inferring the correlation among detected contents. Furthermore, we discover for the first time that the complexity of data has an impact on the training efficiency and subsequent performance of audiovisual model, i.e., more complex data brings more obstacles to the model training, and degrades the performance of downstream audiovisual tasks. To address the issue of audiovisual learning, we propose a novel heterogeneous audiovisual scene analysis module that trains the model from simple to complex scene. We show that such ordered learning procedure rewards the model the merits of easy training and fast convergence. Meanwhile, our audiovisual model can also provide effective unimodal representation and cross-modal alignment performance. We further deploy the well-trained model into practical audiovisual sound localization and separation tasks. We show that our localization model significantly outperforms existing methods, based on which we show comparable performance in sound separation task by comparison to several related SOTA audiovisual learning methods without referring external visualsupervision.

The task of audiovisual sound source localization has been well studied under constrained scenes, where the audio recordings are clean. However, in real world scenarios, audios are usually contaminated by off screen sound and background noise. They will interfere with the procedure of identifying desired sources and building visual sound connections, making previous studies nonapplicable. In this work, we propose the Interference Eraser (IEr) framework, which tackles the problem of audiovisual sound source localization in the wild. The key idea is to eliminate the interference by redefining and carving discriminative audio representations. Specifically, we observe that the previous practice of learning only a single audio representation is insufficient due to the additive nature of audio signals. We thus extend the audio representation with our Audio Instance Identifier module, which clearly distinguishes sounding instances when audio signals of different volumes are unevenly mixed. Then we erase the influence of the audible but off screen sounds and the silent but visible objects by a Cross modal Referrer module with cross modality distillation. Quantitative and qualitative evaluations demonstrate that our framework achieves superior results on sound localization tasks, especially under real world scenarios.

Multiple modalities can provide rich semantic information; and exploiting such information will normally lead to better performance compared with the single-modality counterpart. However, it is not easy to devise an effective cross-modal fusion structure due to the variations of feature dimensions and semantics, especially when the inputs even come from different sensors, as in the field of audio-visual learning. In this work, we propose SepFusion, a novel framework that can smoothly produce optimal fusion structures for visual-sound separation. The framework is composed of two components, namely the model generator and the evaluator. To construct the generator, we devise a lightweight architecture space that can adapt to different input modalities. In this way, we can easily obtain audio-visual fusion structures according to our demands. For the evaluator, we adopt the idea of neural architecture search to select superior networks effectively. This automatic process can significantly save human efforts while achieving competitive performances. Moreover, since our SepFusion provides a series of strong models, we can utilize the model family for broader applications, such as further promoting performance via model assembly, or providing suitable architectures for the separation of certain instrument classes. These potential applications further enhance the competitiveness of our approach.

In this paper, we focus on the Audio-Visual Question Answering (AVQA) task, which aims to answer questions regarding different visual objects, sounds, and their associations in videos. The problem requires comprehensive multimodal understanding and spatio-temporal reasoning over audio-visual scenes. To benchmark this task and facilitate our study, we introduce a large-scale MUSIC-AVQA dataset, which contains more than 45K question-answer pairs covering 33 different question templates spanning over different modalities and question types. We develop several baselines and introduce a spatio-temporal grounded audio-visual network for the AVQA problem. Our results demonstrate that AVQA benefits from multisensory perception and our model outperforms recent A-, V-, and AVQA approaches. We believe that our built dataset has the potential to serve as testbed for evaluating and promoting progress in audio-visual scene understanding and spatio-temporal reasoning. Code and dataset: http://gewu-lab.github.io/MUSIC-AVQA/

Multimodal learning helps to comprehensively understand the world, by integrating different senses. Accordingly, multiple input modalities are expected to boost model performance, but we actually find that they are not fully exploited even when the multimodal model outperforms its uni-modal counterpart. Specifically, in this paper we point out that existing multimodal discriminative models, in which uniform objective is designed for all modalities, could remain under-optimized uni-modal representations, caused by another dominated modality in some scenarios, e.g., sound in blowing wind event, vision in drawing picture event, etc. To alleviate this optimization imbalance, we propose on-the-fly gradient modulation to adaptively control the optimization of each modality, via monitoring the discrepancy of their contribution towards the learning objective. Further, an extra Gaussian noise that changes dynamically is introduced to avoid possible generalization drop caused by gradient modulation. As a result, we achieve considerable improvement over common fusion methods on different multimodal tasks, and this simple strategy can also boost existing multimodal methods, which illustrates its efficacy and versatility. The source code is available at https://github.com/GeWu-Lab/OGM-GE_CVPR2022.

Audiovisual scenes are pervasive in our daily life. It is commonplace for humans to discriminatively localize different sounding objects but quite challenging for machines to achieve class-aware sounding objects localization without category annotations, i.e., localizing the sounding object and recognizing its category. To address this problem, we propose a two-stage step-by-step learning framework to localize and recognize sounding objects in complex audiovisual scenarios using only the correspondence between audio and vision. First, we propose to determine the sounding area via coarse-grained audiovisual correspondence in the single source cases. Then visual features in the sounding area are leveraged as candidate object representations to establish a category-representation object dictionary for expressive visual character extraction. We generate class-aware object localization maps in cocktail-party scenarios and use audiovisual correspondence to suppress silent areas by referring to this dictionary. Finally, we employ category-level audiovisual consistency as the supervision to achieve fine-grained audio and sounding object distribution alignment. Experiments on both realistic and synthesized videos show that our model is superior in localizing and recognizing objects as well as filtering out silent ones. We also transfer the learned audiovisual network into the unsupervised object detection task, obtaining reasonable performance.

Temporal relational modeling in video is essential for human action understanding, such as action recognition and action segmentation. Although Graph Convolution Networks (GCNs) have shown promising advantages in relation reasoning on many tasks, it is still a challenge to apply graph convolution networks on long video sequences effectively. The main reason is that large number of nodes (i.e., video frames) makes GCNs hard to capture and model temporal relations in videos. To tackle this problem, in this paper, we introduce an effective GCN module, Dilated Temporal Graph Reasoning Module (DTGRM), designed to model temporal relations and dependencies between video frames at various time spans. In particular, we capture and model temporal relations via constructing multi-level dilated temporal graphs where the nodes represent frames from different moments in video. Moreover, to enhance temporal reasoning ability of the proposed model, an auxiliary self-supervised task is proposed to encourage the dilated temporal graph reasoning module to find and correct wrong temporal relations in videos. Our DTGRM model outperforms state-of-the-art action segmentation models on three challenging datasets: 50Salads, Georgia Tech Egocentric Activities (GTEA), and the Breakfast dataset. The code is available at https://github.com/redwang/DTGRM.

Unsupervised domain adaptation (UDA) methods for person re-identification (re-ID) aim at transferring re-ID knowledge from labeled source data to unlabeled target data. Although achieving great success, most of them only use limited data from a single-source domain for model pre-training, making the rich labeled data insufficiently exploited. To make full use of the valuable labeled data, we introduce the multi-source concept into UDA person re-ID field, where multiple source datasets are used during training. However, because of domain gaps, simply combining different datasets only brings limited improvement. In this paper, we try to address this problem from two perspectives, i.e. domain-specific view and domain-fusion view. Two constructive modules are proposed, and they are compatible with each other. First, a rectification domain-specific batch normalization (RDSBN) module is explored to simultaneously reduce domain-specific characteristics and increase the distinctiveness of person features. Second, a graph convolutional network (GCN) based multi-domain information fusion (MDIF) module is developed, which minimizes domain distances by fusing features of different domains. The proposed method outperforms state-of-the-art UDA person re-ID methods by a large margin, and even achieves comparable performance to the supervised approaches without any post-processing techniques.

There are rich synchronized audio and visual events in our daily life. Inside the events, audio scenes are associated with the corresponding visual objects; meanwhile, sounding objects can indicate and help to separate their individual sounds in the audio track. Based on this observation, in this paper, we propose a cyclic co-learning (CCoL) paradigm that can jointly learn sounding object visual grounding and audio-visual sound separation in a unified framework. Concretely, we can leverage grounded object-sound relations to improve the results of sound separation. Meanwhile, benefiting from discriminative information from separated sounds, we improve training example sampling for sounding object grounding, which builds a co-learning cycle for the two tasks and makes them mutually beneficial. Extensive experiments show that the proposed framework outperforms the compared recent approaches on both tasks, and they can benefit from each other with our cyclic co-learning. The source code and pre-trained models are released in https://github.com/YapengTian/CCOL-CVPR21.

Discriminatively localizing sounding objects in cocktail-party, i.e., mixed sound scenes, is commonplace for humans, but still challenging for machines. In this paper, we propose a two-stage learning framework to perform self-supervised class-aware sounding object localization. First, we propose to learn robust object representations by aggregating the candidate sound localization results in the single source scenes. Then, class-aware object localization maps are generated in the cocktail-party scenarios by referring the pre-learned object knowledge, and the sounding objects are accordingly selected by matching audio and visual object category distributions, where the audiovisual consistency is viewed as the self-supervised signal. Experimental results in both realistic and synthesized cocktail-party videos demonstrate that our model is superior in filtering out silent objects and pointing out the location of sounding objects of different classes. Code is available at https://github.com/DTaoo/Discriminative-Sounding-Objects-Localization.

Visual-to-auditory sensory substitution devices can assist the blind in sensing the visual environment by translating the visual information into a sound pattern. To improve the translation quality, the task performances of the blind are usually employed to evaluate different encoding schemes. In contrast to the toilsome human-based assessment, we argue that machine model can be also developed for evaluation, and more efficient. To this end, we firstly propose two distinct cross-modal perception model w.r.t. the late-blind and congenitally-blind cases, which aim to generate concrete visual contents based on the translated sound. To validate the functionality of proposed models, two novel optimization strategies w.r.t. the primary encoding scheme are presented. Further, we conduct sets of human-based experiments to evaluate and compare them with the conducted machine-based assessments in the cross-modal generation task. Their highly consistent results w.r.t. different encoding schemes indicate that using machine model to accelerate optimization evaluation and reduce experimental cost is feasible to some extent, which could dramatically promote the upgrading of encoding scheme then help the blind to improve their visual perception ability.

The seen birds twitter, the running cars accompany with noise, etc. These naturally audiovisual correspondences provide the possibilities to explore and understand the outside world. However, the mixed multiple objects and sounds make it intractable to perform efficient matching in the unconstrained environment. To settle this problem, we propose to adequately excavate audio and visual components and perform elaborate correspondence learning among them. Concretely, a novel unsupervised audiovisual learning model is proposed, named as Deep Multimodal Clustering (DMC),that synchronously performs sets of clustering with multimodal vectors of convolutional maps in different shared spaces for capturing multiple audiovisual correspondences. And such integrated multimodal clustering network can be effectively trained with max-margin loss in the end-to-end fashion. Amounts of experiments in feature evaluation and audiovisual tasks are performed. The results demonstrate that DMC can learn effective unimodal representation, with which the classifier can even outperform human performance. Further, DMC shows noticeable performance in sound localization, multisource detection, and audiovisual understanding.

Multiple modalities can provide more valuable information than single one by describing the same contents in various ways. Previous methods mainly focus on fusing the shallow features or high-level representations generated by unimodal deep networks, which only capture part of the hierarchical correlations across modalities. In this paper, we propose to densely integrate the representations by greedily stacking multiple shared layers between different modality-specific networks, which is named as Dense Multimodal Fusion (DMF). The joint representations in different shared layers can capture the correlations in different levels, and the connection between shared layers also provides an efficient way to learn the dependence among hierarchical correlations. These two properties jointly contribute to the multiple learning paths in DMF, which results in faster convergence, lower training loss, and better performance. We evaluate our model on audiovisual speech recognition and cross-modal retrieval. The noticeable performance demonstrates that our model can learn more effective joint representation.

To meet the required huge data analysis, organization, and storage demand, the hashing technique has got a lot of attention as it aims to learn an efficient binary representation from the original high-dimensional data. In this paper, we focus on the unsupervised spectral hashing due to its effective manifold embedding. Existing spectral hashing methods mainly suffer from two problems, i.e., the inefficient spectral candidate and intractable binary constraint for spectral analysis. To overcome these two problems, we propose to employ spectral rotation to seek a better spectral solution and adopt the alternating projection algorithm to settle the complex code constraints, which are therefore named as Spectral Hashing with Spectral Rotation and Alternating Discrete Spectral Hashing, respectively. To enjoy the merits of both methods, the spectral rotation technique is finally combined with the original spectral objective, which aims to simultaneously learn better spectral solution and more efficient discrete codes and is called as Discrete Spectral Hashing. Furthermore, the efficient optimization algorithms are also provided, which just take comparable time complexity to existing hashing methods. To evaluate the proposed three methods, extensive comparison experiments and studies are conducted on four large-scale data sets for the image retrieval task, and the noticeable performance beats several state-of-the-art spectral hashing methods on different evaluation metrics.

To satisfy the huge storage space and organization capacity requirements in addressing big multimodal data, hashing techniques have been widely employed to learn binary representations in cross-modal retrieval tasks. However, optimizing the hashing objective under the necessary binary constraint is truly a difficult problem. A common strategy is to relax the constraint and perform individual binarizations over the learned real-valued representations. In this paper, in contrast to conventional two-stage methods, we propose to directly learn the binary codes, where the model can be easily optimized by a standard gradient descent optimizer. However, before that, we present a theoretical guarantee of the effectiveness of the multimodal network in preserving the inter- and intra-modal consistencies. Based on this guarantee, a novel multimodal deep binary reconstruction model is proposed, which can be trained to simultaneously model the correlation across modalities and learn the binary hashing codes. To generate binary codes and to avoid the tiny gradient problem, a novel activation function first scales the input activations to suitable scopes and, then, feeds them to the tanh function to build the hashing layer. Such a composite function is named adaptive tanh. Both linear and nonlinear scaling methods are proposed and shown to generate efficient codes after training the network. Extensive ablation studies and comparison experiments are conducted for the image2text and text2image retrieval tasks; the method is found to outperform several state-of-the-art deep-learning methods with respect to different evaluation metrics.

Faced with the requirements of huge amounts of data processing nowadays, hashing techniques have attracted much attention due to their efficient storage and searching ability. Among these techniques, the ones based on spectral graph show remarkable performance as they could embed the data on a low-dimensional manifold and maintain the neighborhood structure via a non-linear spectral eigenmap. However, the spectral solution in real value of such methods may deviate from the discrete solution. The common practice is just performing a simple rounding operation to obtain the final binary codes, which could break constraints and even result in worse condition. In this paper, we propose to impose a so-called spectral rotation technique to the spectral hashing objective, which could transform the candidate solution into a new one that better approximates the discrete one. Moreover, the binary codes are obtained from the modified solution via minimizing the Euclidean distance, which could result in more semantical correlation within the manifold, where the constraints for codes are always held. We provide an efficient alternative algorithm to solve the above problems. And a manifold learning perceptive for motivating the proposed method is also shown. Extensive experiments are conducted on three large-scale benchmark datasets and the results show our method outperforms state-of-the-art hashing methods, especially the spectral graph ones.

Image is usually taken for expressing some kinds of emotions or purposes, such as love, celebrating Christmas. There is another better way that combines the image and relevant song to amplify the expression, which has drawn much attention in the social network recently. Hence, the automatic selection of songs should be expected. In this paper, we propose to retrieve semantic relevant songs just by an image query, which is named as the image2song problem. Motivated by the requirements of establishing correlation in semantic/content, we build a semantic-based song retrieval framework, which learns the correlation between image content and lyric words. This model uses a convolutional neural network to generate rich tags from image regions, a recurrent neural network to model lyric, and then establishes correlation via a multi-layer perceptron. To reduce the content gap between image and lyric, we propose to make the lyric modeling focus on the main image content via a tag attention. We collect a dataset from the social-sharing multimodal data to study the proposed problem, which consists of (image, music clip, lyric) triplets. We demonstrate that our proposed model shows noticeable results in the image2song retrieval task and provides suitable songs. Besides, the song2image task is also performed.

With the increasing demand of massive multimodal data storage and organization, cross-modal retrieval based on hashing technique has drawn much attention nowadays. It takes the binary codes of one modality as the query to retrieve the relevant hashing codes of another modality. However, the existing binary constraint makes it difficult to find the optimal cross-modal hashing function. Most approaches choose to relax the constraint and perform thresholding strategy on the real-value representation instead of directly solving the original objective. In this paper, we first provide a concrete analysis about the effectiveness of multimodal networks in preserving the inter- and intra-modal consistency. Based on the analysis, we provide a so-called Deep Binary Reconstruction (DBRC) network that can directly learn the binary hashing codes in an unsupervised fashion. The superiority comes from a proposed simple but efficient activation function, named as Adaptive Tanh (ATanh). The ATanh function can adaptively learn the binary codes and be trained via back-propagation. Extensive experiments on three benchmark datasets demonstrate that DBRC outperforms several state-of-the-art methods in both image2text and text2image retrieval task.

In view of the advantages of deep networks in producing useful representation, the generated features of different modality data (such as image, audio) can be jointly learned using Multimodal Restricted Boltzmann Machines (MRB-M). Recently, audiovisual speech recognition based the M-RBM has attracted much attention, and the MRBM shows its effectiveness in learning the joint representation across audiovisual modalities. However, the built networks have weakness in modeling the multimodal sequence which is the natural property of speech signal. In this paper, we will introduce a novel temporal multimodal deep learning architecture, named as Recurrent Temporal Multimodal RB-M (RTMRBM), that models multimodal sequences by transforming the sequence of connected MRBMs into a probabilistic series model. Compared with existing multimodal networks, it's simple and efficient in learning temporal joint representation. We evaluate our model on audiovisual speech datasets, two public (AVLetters and AVLetters2) and one self-build. The experimental results demonstrate that our approach can obviously improve the accuracy of recognition compared with standard MRBM and the temporal model based on conditional RBM. In addition, RTMRBM still outperforms non-temporal multimodal deep networks in the presence of the weakness of long-term dependencies.

Deep learning is skilled at learning representation from raw data, which are embedded in the semantic space. Traditional multimodal networks take advantage of this, and maximize the joint distribution over the representations of different modalities. However, the similarity among the representations are not emphasized, which is an important property for multimodal data. In this paper, we will introduce a novel learning method for multimodal networks, named as Semantic Similarity Learning (SSL), which aims at training the model via enhancing the similarity between the high-level features of different modalities. Sets of experiments are conducted for evaluating the method on different multimodal networks and multiple tasks. The experimental results demonstrate the effectiveness of SSL in keeping the shared information and improving the discrimination. Particularly, SSL shows its ability in encouraging each modality to learn transferred knowledge from the other one when faced with missing data.