ACM MM 2024 Papers

Model compression and distillation techniques have become essential for deploying deep learning models efficiently. However, existing methods often encounter challenges related to model generalization and scalability for harnessing the expertise of pre-trained large models. This paper introduces CoTuning, a novel framework designed to enhance the generalization ability of neural networks by leveraging collaborative learning between large and small models. CoTuning overcomes the limitations of traditional compression and distillation techniques by introducing strategies for knowledge exchange and simultaneous optimization. Our framework comprises an adapter-based co-tuning mechanism between cloud and edge models, a scale-shift projection for feature alignment, and a novel collaborative knowledge distillation mechanism for domain-agnostic tasks. Extensive experiments conducted on various benchmark datasets demonstrate the effectiveness of CoTuning in improving model generalization while maintaining computational efficiency and scalability. The proposed framework exhibits a significant advancement in model compression and distillation, with broad implications for research in the collaborative evolution of large-small models.

Motion transitions, which serve as bridges between two sequences of character animation, play a crucial role in creating long variable animation for real-time 3D interactive applications. In this paper, we present a framework to produce hybrid character animation, which combines motion capture animation and physical simulation animation that seamlessly connects the front and back motion clips. In contrast to previous works using interpolation for transition, our physics-based approach inherently ensures physical validity, and both the transition moment of the source motion clip and the horizontal rotation of the target motion clip can be specified arbitrarily within a certain range, which achieves high responsiveness and wide latitude for user control. The control policy of character can be trained automatically using only the motion capture data that requires transition, and is enhanced by our proposed Self-Behavior Cloning (SBC), an approach to improve the unsupervised reinforcement learning of motion transition. We show that our framework can accomplish the interactive transition tasks from a fully-connected state machine constructed from nine motion clips with high accuracy and naturalness.

Existing open-vocabulary object detectors require an accurate and compact vocabulary pre-defined during inference. Their performance is largely degraded in real scenarios where the underlying vocabulary may be indeterminate and often exponentially large. To have a more comprehensive understanding of this phenomenon, we propose a new setting called Large-and-Open Vocabulary object Detection, which simulates real scenarios by testing detectors with large vocabularies containing thousands of unseen categories. The vast unseen categories inevitably lead to an increase in category distractors, severely impeding the recognition process and leading to unsatisfactory detection results. To address this challenge, We propose a Large and Open Vocabulary Detector (LOVD) with two core components, termed the Image-to-Region Filtering (IRF) module and Cross-View Verification (CV2) scheme. To relieve the category distractors of the given large vocabularies, IRF performs image-level recognition to build a compact vocabulary relevant to the image scene out of the large input vocabulary, followed by region-level classification upon the compact vocabulary. CV2 further enhances the IRF by conducting image-to-region filtering in both global and local views and produces the final detection categories through a two-branch voting mechanism. Compared to the prior works, our LOVD is more scalable and robust to large input vocabularies, and can be seamlessly integrated with predominant detection methods to improve their open-vocabulary performance. The code can be found at https://github.com/Altria-luo/LOVD.

Multimodal Emotion Recognition in Conversations (ERC) is a typical multimodal learning task in exploiting various data modalities concurrently. Prior studies on effective multimodal ERC encounter challenges in addressing modality imbalances and optimizing learning across modalities. Dealing with these problems, we present a novel framework named Ada2I, which consists of two inseparable modules namely Adaptive Feature Weighting (AFW) and Adaptive Modality Weighting (AMW) for feature-level and modality-level balancing respectively via leveraging both Inter- and Intra-modal interactions. Additionally, we introduce a refined disparity ratio as part of our training optimization strategy, a simple yet effective measure to assess the overall discrepancy of the model's learning process when handling multiple modalities simultaneously. Experimental results validate the effectiveness of Ada2I with state-of-the-art performance compared to baselines on three benchmark datasets, particularly in addressing modality imbalances.

Recent advances in vision-language pre-trained models like CLIP have greatly enhanced general domain image-text retrieval performance. This success has led scholars to develop methods for applying CLIP to Specific Domain Image-Text Retrieval (SDITR) tasks such as Remote Sensing Image-Text Retrieval (RSITR) and Text-Image Person Re-identification (TIReID). However, these methods for SDITR often neglect two critical aspects: the enhancement of modal-level distribution consistency within the retrieval space and the reduction of CLIP's computational cost during inference. To address these issues, this paper presents a novel framework, Accurate and lightweight learning for specific domain Image-text Retrieval (AIR), based on the CLIP. AIR incorporates a Modal-Level distribution Consistency Enhancement regularization (MLCE) loss and a Self-Pruning Distillation Strategy (SPDS) to improve retrieval precision and computational efficiency. The MLCE loss harmonizes the sample distance distributions within image and text modalities, fostering a retrieval space closer to the ideal state. SPDS employs a strategic knowledge distillation process to transfer deep multimodal insights from CLIP to a shallower level, maintaining only the essential layers for inference, thus achieving model light-weighting. Comprehensive experiments across various datasets in RSITR and TIReID reveal that MLCE loss secures optimal retrieval, while SPDS achieves a favorable balance between accuracy and computational demand during testing.

Emotion recognition aims to discern the emotional state of subjects within an image, relying on subject-centric and contextual visual cues. Current approaches typically follow a two-stage pipeline: first localize subjects by off-the-shelf detectors, then perform emotion classification through the late fusion of subject and context features. However, the complicated paradigm suffers from disjoint training stages and limited fine-grained interaction between subject-context elements. To address the challenge, we present a single-stage emotion recognition approach, employing a Decoupled Subject-Context Transformer (DSCT), for simultaneous subject localization and emotion classification. Rather than compartmentalizing training stages, we jointly leverage box and emotion signals as supervision to enrich subject-centric feature learning. Furthermore, we introduce DSCT to facilitate interactions between fine-grained subject-context cues in a ''decouple-then-fuse'' manner. The decoupled query tokens-subject queries and context queries-gradually intertwine across layers within DSCT, during which spatial and semantic relations are exploited and aggregated. We evaluate our single-stage framework on two widely used context-aware emotion recognition datasets, CAER-S and EMOTIC. Our approach surpasses two-stage alternatives with fewer parameter numbers, achieving a 3.39% accuracy improvement and a 6.46% average precision gain on CAER-S and EMOTIC datasets, respectively. Code and models are available at: https://github.com/Sampson-Lee/DSCT.

Multi-focus image fusion (MFIF) aims to combine multiple images with different focused regions into a single all-in-focus image. Existing unsupervised deep learning-based methods only fuse structural information of images in the spatial domain, neglecting potential solutions from the frequency domain exploration. In this paper, we make the first attempt to integrate spatial-frequency information to achieve high-quality MFIF. We propose a novel unsupervised spatial-frequency interaction MFIF network named SFIMFN, which consists of three key components: Adaptive Frequency Domain Information Interaction Module (AFIM), Ret-Attention-Based Spatial Information Extraction Module (RASEM), and Invertible Dual-domain Feature Fusion Module (IDFM). Specifically, in AFIM, we interactively explore global contextual information by combining the amplitude and phase information of multiple images separately. In RASEM, we design a customized transformer to encourage the network to capture important local high-frequency information by redesigning the self-attention mechanism with a bidirectional, two-dimensional form of explicit decay. Finally, we employ IDFM to fuse spatial-frequency information without information loss to generate the desired all-in-focus image. Extensive experiments on different datasets demonstrate that our method significantly outperforms state-of-the-art unsupervised methods in terms of qualitative and quantitative metrics as well as the generalization ability.

Despite the growing prevalence of black-box pre-trained models (PTMs) such as prediction API services, there remains a significant challenge in directly applying general models to real-world scenarios due to the data distribution gap. Considering a data deficiency and constrained computational resource scenario, this paper proposes a novel parameter-efficient transfer learning framework for vision recognition models in the black-box setting. Our framework incorporates two novel training techniques. First, we align the input space (i.e., image) of PTMs to the target data distribution by generating visual prompts of spatial and frequency domain. Along with the novel spatial-frequency hybrid visual prompter, we design a novel training technique based on probabilistic clusters, which can enhance class separation in the output space (i.e., prediction probabilities). In experiments, our model demonstrates superior performance in a few-shot transfer learning setting across extensive visual recognition datasets, surpassing state-of-the-art baselines. Additionally, we show that the proposed method efficiently reduces computational costs for training and inference phases.

Language-based image colorization aims to convert grayscale images to plausible and visually pleasing color images with language guidance, enjoying wide applications in historical photo restoration and film industry. Existing methods mainly leverage large language models and diffusion models to incorporate language guidance into the colorization process. However, it is still a great challenge to build accurate correspondence between the gray image and the semantic instructions, leading to mismatched, overflowing and under-saturated colors. In this paper, we introduce a novel coarse-to-fine framework, COlorfulness COntrollable Language-based Colorization (COCO-LC), that effectively reinforces the image-text correspondence with a coarsely colorized results. In addition, a multi-level condition that leverages both low-level and high-level cues of the gray image is introduced to realize accurate semantic-aware colorization without color overflows. Furthermore, we condition COCO-LC with a scale factor to determine the colorfulness of the output, flexibly meeting the different needs of users. We validate the superiority of COCO-LC over state-of-the-art image colorization methods in accurate, realistic and controllable colorization through extensive experiments. The code and demo will be released at https://lyf1212.github.io/COCO-LC.

News Captioning involves generating the descriptions for news images based on the detailed content of related news articles. Given that these articles often contain extensive information not directly related to the image, captions may end up misaligned with the visual content. To mitigate this issue, we propose the novel cross-modal coherence-enhanced feedback prompting method to clarify the crucial elements that align closely with the visual content for news captioning. Specifically, we first adapt CLIP to develop a news-specific image-text matching module, enriched with insights from language model MPNet using a matching-score comparative loss, which facilitates effective cross-modal knowledge distillation. This module enhance the coherence between images and each news sentences via rating confidence. Then, we design confidence-aware prompts to fine-tune LLaVA model with by LoRa strategy, focusing on essential details in extensive articles. Lastly, we evaluate the generated news caption with refined CLIP, constructing confidence-feedback prompts to further enhance LLaVA through feedback learning, which iteratively refine captions to improve its accuracy. Extensive experiments conduct on two public datasets, GoodNews and NYTimes800k, have validated the effectiveness of our method.

In various domains such as transportation, resource management, and weather forecasting, there is an urgent need for methods that can provide predictions over a sufficiently long time horizon to encompass the period required for decision-making and implementation. Compared to traditional time series forecasting, ultra-long time series forecasting requires enhancing the model's ability to infer long time series, while maintaining inference costs within an acceptable range. To address this challenge, we propose the Boundary-Aware Periodicity-based sparsification strategy for Ultra-Long time series forecasting (BAP-UL).This method effectively captures periodic features in time series and reorganizes inputs and outputs into shorter sub-sequences for improved prediction accuracy. In the paper, we investigate several commonly used benchmark datasets and demonstrate that the proposed method can yield comparable performance across them.

Previous works have shown that reducing parameter overhead and computations for transformer-based single image super-resolution (SISR) models (e.g., SwinIR) usually leads to a reduction of performance. In this paper, we present GRFormer, an efficient and lightweight method, which not only reduces the parameter overhead and computations, but also greatly improves performance. The core of GRFormer is Grouped Residual Self-Attention (GRSA), which is specifically oriented towards two fundamental components. Firstly, it introduces a novel grouped residual layer (GRL) to replace the Query, Key, Value (QKV) linear layer in self-attention, aimed at efficiently reducing parameter overhead, computations, and performance loss at the same time. Secondly, it integrates a compact Exponential-Space Relative Position Bias (ES-RPB) as a substitute for the original relative position bias to improve the ability to represent position information while further minimizing the parameter count. Extensive experimental results demonstrate that GRFormer outperforms state-of-the-art transformer-based methods for ×2, ×3 and ×4 SISR tasks, notably outperforming SOTA by a maximum PSNR of 0.23dB when trained on the DIV2K dataset, while reducing the number of parameter and MACs by about 60% and 49% in only self-attention module respectively. We hope that our simple and effective method that can easily applied to SR models based on window-division self-attention can serve as a useful tool for further research in image super-resolution. The code is available at https://github.com/sisrformer/GRFormer.

This work proposes a novel and simple sequential learning strategy to train models on videos and texts for multimodal sentiment analysis. To estimate sentiment polarities on unseen out-of-distribution data, we introduce a multimodal model that is trained either in a single source domain or multiple source domains using our learning strategy. This strategy starts with learning domain invariant features from text, followed by learning sparse domain-agnostic features from videos, assisted by the selected features learned in text. Our experimental results demonstrate that our model achieves significantly better performance than the state-of-the-art approaches on average in both single-source and multi-source settings. Our feature selection procedure favors the features that are independent to each other and are strongly correlated with their polarity labels. To facilitate research on this topic, the source code of this work will be publicly available upon acceptance.

Sign language recognition (SLR) refers to interpreting sign language glosses from given videos automatically. This research area presents a complex challenge in computer vision because of the rapid and intricate movements inherent in sign languages, which encompass hand gestures, body postures, and even facial expressions. Recently, skeleton-based action recognition has attracted increasing attention due to its ability to handle variations in subjects and backgrounds independently. However, current skeleton-based SLR methods exhibit three limitations: 1) they often neglect the importance of realistic hand poses, where most studies train SLR models on non-realistic skeletal representations; 2) they tend to assume complete data availability in both training or inference phases, and capture intricate relationships among different body parts collectively; 3) these methods treat all sign glosses uniformly, failing to account for differences in complexity levels regarding skeletal representations. To enhance the realism of hand skeletal representations, we present a kinematic hand pose rectification method for enforcing constraints. Mitigating the impact of missing data, we propose a feature-isolated mechanism to focus on capturing local spatial-temporal context. This method captures the context concurrently and independently from individual features, thus enhancing the robustness of the SLR model. Additionally, to adapt to varying complexity levels of sign glosses, we develop an input-adaptive inference approach to optimise computational efficiency and accuracy. Experimental results demonstrate the effectiveness of our approach, as evidenced by achieving a new state-of-the-art (SOTA) performance on WLASL100 and LSA64. For WLASL100, we achieve a top-1 accuracy of 86.50\%, marking a relative improvement of 2.39% over the previous SOTA. For LSA64, we achieve a top-1 accuracy of 99.84%.

In the realm of cross-modal retrieval, seamlessly integrating diverse modalities within multimedia remains a formidable challenge, especially given the complexities introduced by noisy correspondence learning (NCL). Such noise often stems from mismatched data pairs, which is a significant obstacle distinct from traditional noisy labels. This paper introduces Pseudo-Classification based Pseudo-Captioning (PC$^2$) framework to address this challenge. PC$^2$ offers a threefold strategy: firstly, it establishes an auxiliary "pseudo-classification" task that interprets captions as categorical labels, steering the model to learn image-text semantic similarity through a non-contrastive mechanism. Secondly, unlike prevailing margin-based techniques, capitalizing on PC$^2$'s pseudo-classification capability, we generate pseudo-captions to provide more informative and tangible supervision for each mismatched pair. Thirdly, the oscillation of pseudo-classification is borrowed to assistant the correction of correspondence. In addition to technical contributions, we develop a realistic NCL dataset called Noise of Web (NoW), which could be a new powerful NCL benchmark where noise exists naturally. Empirical evaluations of PC$^2$ showcase marked improvements over existing state-of-the-art robust cross-modal retrieval techniques on both simulated and realistic datasets with various NCL settings. The contributed dataset and source code are released at https://github.com/alipay/PC2-NoiseofWeb.

Groundbreaking advancements in text-to-image generation have recently been achieved with the emergence of diffusion models. These models exhibit a remarkable ability to generate highly artistic and intricately detailed images based on textual prompts. However, obtaining desired generation outcomes often necessitates repetitive trials of manipulating text prompts just like casting spells on a magic mirror, and the reason behind that is the limited capability of semantic understanding inherent in current image generation models. Specifically, existing diffusion models encode the text prompt input with a pre-trained encoder structure, which is usually trained on a limited number of image-caption pairs. The state-of-the-art large language models (LLMs) based on the decoder-only structure have shown a powerful semantic understanding capability as their architectures are more suitable for training on very large-scale unlabeled data. In this work, we propose to enhance text-to-image diffusion models by borrowing the strength of semantic understanding from large language models, and devise a simple yet effective adapter to allow the diffusion models to be compatible with the decoder-only structure. Meanwhile, we also provide a supporting theoretical analysis with various architectures (e.g., encoder-only, encoder-decoder, and decoder-only), and conduct extensive empirical evaluations to verify its effectiveness. The experimental results show that the enhanced models with our adapter module are superior to the stat-of-the-art models in terms of text-to-image generation quality and reliability.

In this paper, we present the One-shot In-context Part Segmentation (OIParts) framework, designed to tackle the challenges of part segmentation by leveraging visual foundation models (VFMs). Existing training-based one-shot part segmentation methods that utilize VFMs encounter difficulties when faced with scenarios where the one-shot image and test image exhibit significant variance in appearance and perspective, or when the object in the test image is partially visible. We argue that training on the one-shot example often leads to overfitting, thereby compromising the model's generalization capability. Our framework offers a novel approach to part segmentation that is training-free, flexible, and data-efficient, requiring only a single in-context example for precise segmentation with superior generalization ability. By thoroughly exploring the complementary strengths of VFMs, specifically DINOv2 and Stable Diffusion, we introduce an adaptive channel selection approach by minimizing the intra-class distance for better exploiting these two features, thereby enhancing the discriminatory power of the extracted features for the fine-grained parts. We have achieved remarkable segmentation performance across diverse object categories. The OIParts framework not only eliminates the need for extensive labeled data but also demonstrates superior generalization ability. Through comprehensive experimentation on three benchmark datasets, we have demonstrated the superiority of our proposed method over existing part segmentation approaches in one-shot settings.