Large language models (LLMs) have advanced to encompass extensive knowledge across diverse domains. Yet controlling what a LLMs should not know is important for ensuring alignment and thus safe use. However, effective unlearning in LLMs is difficult due to the fuzzy boundary between knowledge retention and forgetting. This challenge is exacerbated by entangled parameter spaces from continuous multi-domain training, often resulting in collateral damage, especially under aggressive unlearning strategies. Furthermore, the computational overhead required to optimize State-of-the-Art (SOTA) models with billions of parameters poses an additional barrier. In this work, we present ALTER, a lightweight unlearning framework for LLMs to address both the challenges of knowledge entanglement and unlearning efficiency. ALTER operates through two phases: (I) high entropy tokens are captured and learned via the shared A matrix in LoRA, followed by (II) an asymmetric LoRA architecture that achieves a specified forgetting objective by parameter isolation and unlearning tokens within the target subdomains. Serving as a new research direction for achieving unlearning via token-level isolation in the asymmetric framework. ALTER achieves SOTA performance on TOFU, WMDP, and MUSE benchmarks with over 95% forget quality and shows minimal side effects through preserving foundational tokens. By decoupling unlearning from LLMs' billion-scale parameters, this framework delivers excellent efficiency while preserving over 90% of model utility, exceeding baseline preservation rates of 47.8-83.6%.

Multimodal Large Language Models (MLLMs) with unified architectures excel across a wide range of vision-language tasks, yet aligning them with personalized image generation remains a significant challenge. Existing methods for MLLMs are frequently subject-specific, demanding a data-intensive fine-tuning process for every new subject, which limits their scalability. In this paper, we introduce MM-R1, a framework that integrates a cross-modal Chain-of-Thought (X-CoT) reasoning strategy to unlock the inherent potential of unified MLLMs for personalized image generation. Specifically, we structure personalization as an integrated visual reasoning and generation process: (1) grounding subject concepts by interpreting and understanding user-provided images and contextual cues, and (2) generating personalized images conditioned on both the extracted subject representations and user prompts. To further enhance the reasoning capability, we adopt Grouped Reward Proximal Policy Optimization(GRPO) to explicitly align the generation. Experiments demonstrate that MM-R1 unleashes the personalization capability of unified MLLMs to generate images with high subject fidelity and strong text alignment in a zero-shot manner.

Fully fine-tuning large pre-trained models for each downstream task is impractical due to prohibitive memory, computation, and storage costs. Although parameter-efficient fine-tuning (PEFT) methods address this issue, leading methods like LoRA still exhibit linear scaling of trainable parameters with hidden size. Recent studies have explored PEFT in the frequency domain to reduce computational costs by employing fast Fourier transform and discrete cosine transform with sparse frequency selection. These methods rely on global frequency representations that lack spatial locality and disperse energy across the domain. As a result, sparse coefficient selection struggles to preserve fine-grained structural information and often introduces artifacts such as ringing near boundaries. To address these limitations, we propose DWTSG, a novel PEFT framework based on discrete wavelet transform (DWT) and subband guidance. DWTSG decomposes pre-trained weights into four wavelet subbands that jointly encode global context and local details. It fine-tunes only the most informative coefficients in each subband through an energy-based selection strategy that prioritizes coefficients based on their individual importance and interactions. Finally, inverse DWT reconstructs the updated weights, enabling efficient and precise adaptation. Extensive experiments on natural language understanding, commonsense reasoning, and image classification demonstrate that DWTSG outperforms existing PEFT methods, achieving superior performance and higher parameter efficiency.

Jiaquan Zhang, Chaoning Zhang, Shuxu Chen, Zhenzhen Huang, Pengcheng Zheng, Zhicheng Wang, Ping Guo, Fan Mo, Sung-Ho Bae, Jie Zou, Jiwei Wei, Yang Yang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Session-based recommendation aims to predict the next item that anonymous users may be interested in, based on their current session interactions. Recent studies have demonstrated that retrieving neighbor sessions to augment the current session can effectively alleviate the data sparsity issue and improve recommendation performance. However, existing methods typically rely on explicitly observed session data, neglecting latent neighbors - not directly observed but potentially relevant within the interest space - thereby failing to fully exploit the potential of neighbor sessions in recommendation.

The rapid growth of model scale has necessitated substantial computational resources for fine-tuning. Existing approach such as low-rank adaptation (LoRA) has sought to address the problem of handling the large updated parameters in full fine-tuning (FT). However, LoRA utilize random initialization and optimization of low-rank matrices to approximate updated weights, which can result in suboptimal convergence and an accuracy gap compared to full fine-tuning (FT). To address these issues, we propose low-rank LDU (LoLDU), a parameter-efficient fine-tuning (PEFT) approach that significantly reduces trainable parameters by 2600 times compared to regular PEFT methods while maintaining comparable performance. LoLDU leverages lower-diag-upper (LDU) decomposition to initialize low-rank matrices for faster convergence and nonsingularity. We focus on optimizing the diagonal matrix for scaling transformations. To the best of our knowledge, LoLDU has the fewest parameters among all PEFT approaches. We conducted extensive experiments across 4 instruction-following datasets, six natural language understanding (NLU) datasets, eight image classification datasets, and image generation datasets with multiple model types [LLaMA2, RoBERTa, ViT, and stable diffusion (SD)], providing a comprehensive and detailed analysis. Our open-source code can be accessed at https://anonymous.4open.science/r/LoLDU-B5A6.

Video Moment Retrieval (VMR) involves locating specific moments within a video based on natural language queries. However, existing VMR methods that employ various strategies for cross-modal alignment still face challenges such as limited understanding of fine-grained semantics, semantic overlap, and sparse constraints. To address these limitations, we propose a novel Concept Decomposition Transformer (CDTR) model for VMR. CDTR introduces a semantic concept decomposition module that disentangles video moments and sentence queries into concept representations, reflecting the relevance between various concepts and capturing fine-grained semantics which is crucial for cross-modal matching. These decomposed concept representations are then used as pseudo-labels, determined as positive or negative samples by adaptive concept-specific thresholds. Subsequently, fine-grained concept alignment is performed in video intra-modal and textual-visual cross-modal, aligning different conceptual components within features, enhancing the model's ability to distinguish fine-grained semantics, and alleviating issues related to semantic overlap and sparse constraints. Comprehensive experiments demonstrate the effectiveness of the CDTR, outperforming state-of-the-art methods on three widely used datasets: QVHighlights, Charades-STA, and TACoS.

Generating high-fidelity talking heads that maintain stable head poses and achieve robust lip sync remains a significant challenge. Although methods based on 3D Gaussian Splatting (3DGS) offer a promising solution via point-based deformation, they suffer from inconsistent head dynamics and mismatched mouth movements due to unstable Gaussian initialization and incomplete speech features. To overcome these limitations, we introduce SyncGaussian, a 3DGS-based framework that ensures stable head poses, enhanced lip sync, and realistic appearances with real-time rendering. SyncGaussian employs a stable head Gaussian initialization strategy to mitigate head jitter by optimizing commonly used rough head pose parameters. To enhance lip sync, we propose a sync-enhanced encoder that leverages audio-to-text and audio-to-visual speech features. Guided by a tailored cosine similarity loss function, the encoder integrates discriminative speech features through a multi-level sync adaptation mechanism, enabling the learning of an adaptive speech feature space. Extensive experiments demonstrate that SyncGaussian outperforms state-of-the-art methods in image quality, dynamic motion, and lip sync, with the potential for real-time applications.

Conversational recommender systems aim to provide personalized recommendations by analyzing and utilizing contextual information related to dialogue. However, existing methods typically model the dialogue context as a whole, neglecting the inherent complexity and entanglement within the dialogue. Specifically, a dialogue comprises both focus information and background information, which mutually influence each other. Current methods tend to model these two types of information mixedly, leading to misinterpretation of users' actual needs, thereby lowering the accuracy of recommendations. To address this issue, this paper proposes a novel model to introduce contextual disentanglement for improving conversational recommender systems, named DisenCRS. The proposed model DisenCRS employs a dual disentanglement framework, including self-supervised contrastive disentanglement and counterfactual inference disentanglement, to effectively distinguish focus information and background information from the dialogue context under unsupervised conditions. Moreover, we design an adaptive prompt learning module to automatically select the most suitable prompt based on the specific dialogue context, fully leveraging the power of large language models. Experimental results on two widely used public datasets demonstrate that DisenCRS significantly outperforms existing conversational recommendation models, achieving superior performance on both item recommendation and response generation tasks.

Point clouds serve as the foundational representation of 3D objects, playing a pivotal role in both computer vision and computer graphics. Recently, the acquisition of point clouds has been effortless because of the development of hardware devices. However, the collected point clouds may be incomplete due to environmental conditions, such as occlusion. Therefore, completing partial point clouds becomes an essential task. The majority of current methods address point cloud completion via the utilization of shape priors. While these methods have demonstrated commendable performance, they often encounter challenges in preserving the global structural and geometric details of the 3D shape. In contrast to those mentioned earlier, we propose a novel cross-modal coarse-to-fine network (CMNet) for point cloud completion. Our method utilizes additional image information to provide global information, thus avoiding the loss of structure. To ensure that the generated results contain sufficient geometric details, we propose a coarse-to-fine learning approach based on multiple patches. Specifically, we encode the image and use multiple generators to generate multiple coarse patches, which are combined into a complete shape. Subsequently, based on the coarse patches generated in advance, we generate fine patches by combining partial point cloud information. Experimental results show that our method achieves state-of-the-art performance on point cloud completion.

Video grounding tasks have recently gained significant attention. However, existing methods failed to fully comprehend the semantics within queries and videos, often overlooking key content. Moreover, the lack of fine-grained cross-modal alignment and interaction to guide the semantic matching of complex texts and videos lead to inconsistent representational modeling. To address this issue, we propose a Semantic Hierarchical Grounding model, referred to as SHG, and design a cross-modal semantic hierarchical graph to achieve fine-grained semantic understanding. SHG decomposes both the query and each video moment into three levels: global, action, and element. This topology, ranging from global to local, establishes multigranularity intrinsic connections between the two modalities, fostering a comprehensive understanding of dynamic semantics and fine-grained cross-modal matching. Accordingly, to fully leverage the rich information within the cross-modal semantic hierarchical graph, we employ contrastive learning by seeking samples with the same action and element semantics, then achieve node-moment cross-modal hierarchical matching for global alignment. This approach can unearth fine-grained clues and align semantics across multiple granularities. Moreover, we combine the designed hierarchical graph interaction for coarse-to-fine fusion of text and video, thereby enabling highly accurate video grounding. Extensive experiments conducted on three challenging public datasets (ActivityNet-Captions, TACoS, and Charades-STA) demonstrate that the proposed approach outperforms state-of-the-art techniques, validating its effectiveness.

Temporal Sentence Grounding (TSG) requires a thorough understanding of the complex cross-modal semantic relationships between videos and text. However, existing methods fail to accurately capture content at diverse granularity levels with distinct semantics, making it difficult to achieve fine alignment of visuals and text. To overcome this issue, we attempt to mine for rich semantic clues by utilizing the hierarchical correspondence structure and multi-granularity visual-to-text reconstruction, achieving fine-grained reasoning. Specifically, for the TSG task, we propose a novel Hierarchical Cross-modal Fine-grained Mining Network (HCFMN), which utilizes an attention mechanism based on temporal hierarchical relationships to extract temporal features corresponding to the text of different granularities. We leverage the reconstructability of visual-to-text, recovering multi-granularity textual content from coarse to fine by focusing on temporal features at different layers, hierarchically extracting temporal features and the dependencies related to the text, and implementing fine-grained cross-modal semantic alignment. Furthermore, HCFMN introduces a novel partitioned efficient attention mechanism, which significantly enhances the model’s efficiency through a two-stage attention based on sequence and channel compression. Extensive experimental results on three public datasets (ActivityNet-Captions, TACoS, and Charades-STA) demonstrate that the proposed method achieves state-of-the-art performance.

Point cloud completion aims at completing shapes from their partial. Most existing methods utilized shape’s priors information for point cloud completion, such as inputting the partial and getting the complete one through an encoder-decoder deep learning structure. However, it is very often to easily cause the loss of information in the generation process because of the invisibility of missing areas. Unlike most existing methods directly inferring the missing points using shape priors, we address it as a cross-modality task. We propose a new Cross-modal Dual Phases Network (CDPNet) for shape completion. Our key idea is that the global information of the shape is obtained from the extra single-view image, and the partial point clouds provide the geometric information. After that, the multi-modal features jointly guide the specific structural information. To learn the geometric details of the shape, we chose to use patches to preserve the local geometric feature. In this way, we can generate shapes with enough geometric details. Experimental results show that our method achieves state-of-the-art performance on point cloud completion.

When applying high-level visual algorithms to rainy scenes, it is customary to preprocess the rainy images using low-level rain removal networks, followed by visual networks to achieve the desired objectives. Such a setting has never been explored by adversarial attack methods, which are only limited to attacking one kind of them. Considering the deficiency of multi-functional attacking strategies and the significance for open-world perception scenarios, we are the first to propose a Cascaded Adversarial Attack (CAA) setting, where the adversarial example can simultaneously attack different-level tasks, such as rain removal and semantic segmentation in an integrated system. Specifically, our attack on the rain removal network aims to preserve rain streaks in the output image, while for the semantic segmentation network, we employ powerful existing adversarial attack methods to induce misclassification of the image content. Importantly, CAA innovatively utilizes binary masks to effectively concentrate the aforementioned two significantly disparate perturbation distributions on the input image, enabling attacks on both networks. Additionally, we propose two variants of CAA, which minimize the differences between the two generated perturbations by introducing a carefully designed perturbation interaction mechanism, resulting in enhanced attack performance. Extensive experiments validate the effectiveness of our methods, demonstrating their superior ability to significantly degrade the performance of the downstream task compared to methods that solely attack a single network.

This paper addresses an important and valuable open-world object detection (OWOD) in autonomous driving scenarios, which aims to detect objects under bothdomain-agnosticandcategory-agnosticsettings simultaneously. Existing OWOD algorithms mainly focus on the detection of pre-defined object categories under various conditions (domain-agnostic) or instead perform zero-shot object detection (category-agnostic), separately. The knowledge gap between seen and unseen object categories poses challenges for models optimized with supervision from the only seen object categories. The domain difference across different scenarios also causes further challenges in aligning observations with different appearances. To address these two challenges simultaneously, we propose our Instance Dictionary Learning (IDL for short) for more robust and accurate OWOD performance. We first design a pre-training procedure to build up the mappings between region features and category semantic embeddings by introducing instance contrastive learning. The joint vision-semantic space is formulated through the more detailed instance-level “Dictionary”, which expresses the region-category correspondences and helps link the seen and unseen object categories. The domain discrimination is further designed for extracting the domain invariance feature representations in the further training procedure seamlessly. The proposed IDL could detect the unseen categories from unseen domains without any bounding box annotations while there is no obvious performance drop on detecting seen categories meanwhile. Comprehensive experiments have been conducted and our method could achieve a new state-of-the-art OWOD performance over previous algorithms.

Composed image retrieval aims to search a target image by concurrently understanding the composed inputs with a reference image and the complementary modification text. It aims to find a shared latent space where the representation of the composed inputs is close to the desired target image. Most previous methods capture the one-to-one correspondence between the composed inputs and target image, which encodes the composed inputs and the target image into single points in the feature space. However, the one-to-one correspondence cannot effectively handle this task due to the inherent ambiguity problem arising from the various semantic meanings and data uncertainty. Specifically, the composed inputs and target image always exhibit various semantic meanings, affecting the retrieval results. Moreover, given the composed inputs (resp. target image), there are multiple target images (resp. composed inputs) that equally make sense. In this paper, we propose a novel method termed Set of Diverse Queries with Uncertainty Regularization (SDQUR) to solve such inherent ambiguity problem. First, we utilize diverse queries to adaptively aggregate the composed inputs and target image into multiple deterministic embeddings that capture different semantic meanings in the triplet affecting the retrieval process. It can exploit the deterministic many-to-many correspondence within each triple through these set-based queries. Moreover, we provide an uncertainty regularization module to encode the composed inputs and target image into gaussian distribution. Multiple potential positive candidates are sampled from the distribution for probabilistic many-to-many correspondence. Through the complementary deterministic and probabilistic many-to-many correspondence manner, we achieve consistent improvements on the standard FashionIQ, CIRR, and Shoes benchmarks, surpassing the state-of-the-art methods by a large margin.

Cross-modal retrieval (e.g., query a given image to obtain a semantically similar sentence, and vice versa) is an important but challenging task, as the heterogeneous gap and inconsistent distributions exist between different modalities. The dominant approaches struggle to bridge the heterogeneity by capturing the common representations among heterogeneous data in a constructed subspace which can reflect the semantic closeness. However, insufficient consideration is taken into the fact that learned latent representations are actually heavily entangled with those semantic-unrelated features, which obviously further compounds the challenges of cross-modal retrieval. To alleviate the difficulty, this work makes an assumption that the data are jointly characterized by two independent features: semantic-shared and semantic-unrelated representations. The former presents characteristics of consistent semantics shared by different modalities, while the latter reflects the characteristics with respect to the modality yet unrelated to semantics, such as background, illumination, and other low-level information. Therefore, this paper aims to disentangle the shared semantics from the entangled features, andthus the purer semantic representation can promote the closeness of paired data. Specifically, this paper designs a novel Semantics Disentangling approach for Cross-Modal Retrieval (termed as SDCMR) to explicitly decouple the two different features based on variational auto-encoder. Next, the reconstruction is performed by exchanging shared semantics to ensure the learning of semantic consistency. Moreover, a dual adversarial mechanism is designed to disentangle the two independent features via a pushing-and-pulling strategy. Comprehensive experiments on four widely used datasets demonstrate the effectiveness and superiority of the proposed SDCMR method by achieving a new bar on performance when compared against 15 state-of-the-art methods.

The susceptibility of deep neural networks (DNNs) to adversarial examples has raised significant concerns regarding the security and reliability of artificial intelligence systems. These examples contain maliciously crafted perturbations not perceptible to the human eye but can cause the model to make wrong predictions. Adversarial training (AT) is the de facto standard method for enhancing adversarial robustness. However, the improved robustness is often at the cost of a significant drop in standard accuracy for clean samples. Numerous works have attempted to alleviate this trade-off by identifying its causes. A key factor lies in the variability of clean samples, which leads to different adversarial examples being generated using the same attack strategy. The other factor is the disruption of the underlying data structure caused by adversarial perturbations. To overcome these challenges, we propose a novel adversarial training framework named Hardness-Guided Sample-Dependent Adversarial Training (HGSD-AT), which dynamically adjusts the attack strategy based on the hardness of the current adversarial sample to further improve the robustness of the model. By utilizing the two types of constraints which construct from a temporal perspective and spatial distribution perspective, our method directly learns the impact of attack methods on the model, rather than the indirect effects associated with sample distribution. This approach aims to improve the generation of adversarial examples while simultaneously enhancing the robustness and accuracy of DNNs. Our approach exhibits superior performance in terms of both robustness and natural accuracy compared to state-of-the-art defense methods, as validated through comprehensive experiments conducted on three benchmark datasets.

Multi-view speech emotion recognition (SER) based on the pre-trained model has gained attention in the last two years, which shows great potential in improving the model performance in speaker-independent scenarios. However, the existing work either relies on various fine-tuning methods or uses excessive feature views with complex fusion strategies, causing the increase of complexity with limited performance benefit. In this paper, we improve multi-view SER based on the pre-trained model from the perspective of a low-level speech feature. Specifically, we forgo fine-tuning the pre-trained model and instead focus on learning effective features hidden in the low-level speech feature mel-scale frequency cepstral coefficient (MFCC). We propose atwo-streampoolingchannelattention (TsPCA) module to discriminatively weight the channel dimensions of the features derived from MFCC. This module enables inter-channel interaction and learning of emotion sequence information across channels. Furthermore, we design a simple but effective feature view fusion strategy to learn robust representations. In the comparison experiments, our method achieves the WA and UA of 73.97%/74.69% and 74.61%/75.66% on the IEMOCAP dataset, 97.21% and 97.11% on the Emo-DB dataset, 77.08% and 77.34% on the RAVDESS dataset, and 74.38% and 71.43% on the SAVEE dataset. Extensive experiments on the four datasets demonstrate that our method consistently surpasses existing methods and achieves a new State-of-the-Art result.

The vulnerability of deep neural networks to adversarial examples has raised huge concerns about the security of these algorithms. Black-box adversarial attacks have received a lot of attention as an influential method for evaluating model robustness. While various sophisticated adversarial attack methods have been proposed, the success rate in the black-box scenario still needs to be improved. To address these issues, we develop an Adaptive Multi-scale Degradation-based Attack method calledAMDA. The intuitive motivation behind our approach is that different models tend to have similar attention regions for low-scale images. Specifically, AMDA uses degraded images to generate perturbations at different scales and fuses these perturbations to generate adversarial examples that are insensitive to model changes. Furthermore, we design an adaptive multi-scale perturbation fusion that evaluates the transferability of perturbations at different scales based on noise and adaptively allocates fusion weights to prioritize strong transferability attacks and avoid being compromised by local optima. Extensive experimental results on the ImageNet, CIFAR-100, and CIFAR-10 datasets demonstrate that the proposed AMDA algorithm exhibits competitive performance for both normally trained models and defense models.

Recently, person re-identification has gained significant attention from both academic and industry fields due to its potential applications in surveillance and security. However, the security of re-identification systems has not been widely investigated, and they are vulnerable to adversarial attacks, which can significantly degrade their performance. Although numerous sophisticated adversarial training methods have been proposed for image classification, metric analysis systems such as person re-identification have not been fully explored. In this paper, we develop a novel adversarial training framework with a dynamic attack strategy for person re-identification, to further enhance the robustness of the model. Specifically, we gradually increase the perturbation budget during the generation until the generated adversarial examples reach a certain level of attack strength. As the iterations progress, the model becomes more robust, and our framework can generate stronger adversarial examples to continuously explore the robustness bounds of the model. Moreover, to alleviate the conflict between the adversarial robustness and natural generalization of the model, we design a novel performance alignment loss to further constrain the adversarial example generation process, which can make the generated adversarial examples as close as possible to the clean samples in terms of performance. Experiments on two widely used person re-ID benchmark datasets demonstrate the effectiveness and superiority of our proposed method.

Deep Neural Networks (DNNs) have primarily been demonstrated to be successful when large-scale labeled data are available. However, DNNs usually fail when tasked in few-sample learning scenarios, and the results will be much worse when the limited data show large intra-class variation and inter-class similarity (a.k.a fine-grained classification). To solve this challenging task, the idea of carrying out feature augmentation is visited and better achieved by exploring the merit of the forward Euler method in solving ordinary differential equations (ODEs), and a novel high-order feature augmentation (HFA) model with ResNet is proposed. Specifically, the proposed method leverages the stacked residual structure to model the direction of feature change over the initial state, and uses the triplet loss as constraint to model the step size of change in an adaptive manner. As a result, the initial features can then be augmented by a residual structure with a forward Eulerian form to generate features of the same subcategory with a similar representation as the input image. Furthermore, the proposed augmentation mechanism enjoys two additional benefits: a) it can help avoid the over-fitting issue when learned with insufficient training data; b) it can be used seamlessly with any residual structure-based classification network, and the ResNet used in this paper remains unchanged during testing. Extensive experiments are carried out on fine-grained visual categorization benchmarks, and the results demonstrate that our approach can significantly improve the categorization performance when the training data is highly insufficient.

We study the task of image retrieval with text feedback, where a reference image and modification text are composed to retrieve the desired target image. To accomplish this goal, existing methods always get the multimodal representations through different feature encoders and then adopt different strategies to model the correlation between the composed inputs and the target image. However, the multimodal query brings more challenges as it requires not only the synergistic understanding of the semantics from the heterogeneous multimodal inputs but also the ability to accurately build the underlying semantic correlation existing in each inputs-target triplet, i.e., reference image, modification text, and target image. In this paper, we tackle these issues with a novel Align and Retrieve (AlRet) framework. First, our proposed methods employ the contrastive loss in the feature encoders to learn meaningful multimodal representation while making the subsequent correlation modeling process in a more harmonious space. Then we propose to learn the accurate correlation between the composed inputs and target image in a novel composition-and-decomposition paradigm. Specifically, the composition network couples the reference image and modification text into a joint representation to learn the correlation between the joint representation and target image. The decomposition network conversely decouples the target image into visual and text subspaces to exploit the underlying correlation between the target image with each query element. The composition-and-decomposition paradigm forms a closed loop, which can be optimized simultaneously to promote each other in the performance. Massive comparison experiments on three real-world datasets confirm the effectiveness of the proposed method.

Low-light image enhancement (LLIE) investigates how to improve illumination and produce normal-light images. The majority of existing methods improve low-light images via a global and uniform manner, without taking into account the semantic information of different regions. Without semantic priors, a network may easily deviate from a region's original color. To address this issue, we propose a novel semantic-aware knowledge-guided framework (SKF) that can assist a low-light enhancement model in learning rich and diverse priors encapsulated in a semantic segmentation model. We concentrate on incorporating semantic knowledge from three key aspects: a semantic-aware embedding module that wisely integrates semantic priors in feature representation space, a semantic-guided color histogram loss that preserves color consistency of various instances, and a semantic-guided adversarial loss that produces more natural textures by semantic priors. Our SKF is appealing in acting as a general framework in LLIE task. Extensive experiments show that models equipped with the SKF significantly outperform the baselines on multiple datasets and our SKF generalizes to different models and scenes well. The code is available at Semantic-Aware-Low-Light-Image-Enhancement.

Existing domain adaptive object detection algorithms (DAOD) have demonstrated their effectiveness in discriminating and localizing objects across scenarios. However, these algorithms typically assume a single source and target domain for adaptation, which is not representative of the more complex data distributions in practice. To address this issue, we propose a novel Open-Scenario Domain Adaptive Object Detection (OSDA), which leverages multiple source and target domains for more practical and effective domain adaptation. We are the first to increase the granularity of the background category by building the foundation model using contrastive vision-language pre-training in an open-scenario setting for better distinguishing foreground and background, which is under-explored in previous studies. The performance gains by introducing the pre-training have been observed and have validated the model's ability to detect objects across domains. To further fine-tune the model for domain-specific object detection, we propose a hierarchical feature alignment strategy to obtain a better common feature space among the various source and target domains. In the case of multi-source domains, the cross-reconstruction framework is introduced for learning more domain invariances. The proposed method is able to alleviate knowledge forgetting without any additional computational costs. Extensive experiments across different scenarios demonstrate the effectiveness of the proposed model.

Previous studies on multimodal fake news detection have observed the mismatch between text and images in the fake news and attempted to explore the consistency of multimodal news based on global features of different modalities. However, they fail to investigate this relationship between fine-grained fragments in multimodal content. To gain public trust, fake news often includes relevant parts in the text and the image, making such multimodal content appear consistent. Using global features may suppress potential inconsistencies in irrelevant parts. Therefore, in this paper, we propose a novel Consistency-learning Fine-grained Fusion Network (CFFN) that separately explores the consistency and inconsistency from high-relevant and low-relevant word-region pairs. Specifically, for a multimodal post, we divide word-region pairs into high-relevant and low-relevant parts based on their relevance scores. For the high-relevant part, we follow the cross-modal attention mechanism to explore the consistency. For low-relevant part, we calculate inconsistency scores to capture inconsistent points. Finally, a selection module is used to choose the primary clue (consistency or inconsistency) for identifying the credibility of multimodal news. Extensive experiments on two public datasets demonstrate that our CFFN substantially outperforms all the baselines. Our code can be found at: https://github.com/uestc-lj/CFFN/.

Deep metric learning has become a key component of cross-modal retrieval. By learning to pull the features of matched instances closer while pushing the features of mismatched instances farther away, one can learn highly robust multi-modal representations. Most existing cross-modal retrieval methods leverage vanilla triplet loss to train the network, which cannot adaptively penalize pairs with different hardness. Although various weighting strategies have been designed for unimodal matching tasks, few weighting strategies have been applied to cross-modal tasks due to the specificity of cross-modal tasks. While few weighting strategies are designed for cross-modal scenarios, they usually involve a lot of hyper-parameters, which require a lot of computational resources to fine-tune. In this paper, we introduce a new exponential loss, which can assign appropriate weights to individual positive and negative pairs according to their similarity so that it can adaptively penalize pairs with different hardness. Furthermore, the exponential loss has only two hyper-parameters, making it easier to find the optimal parameters to suit various data distributions in practice. Exponential loss can be universally applied to well-established cross-modal models and further boost their retrieval performance. We exhaustively ablate our method on Image-Text matching, Video-Text matching, as well as unimodal Image matching. Experimental results show that a standard model trained with exponential loss can achieve noticeable performance gains.

In this paper, we study the composed query image retrieval, which aims at retrieving the target image similar to the composed query, i.e., a reference image and the desired modification text. Compared with conventional image retrieval, this task is more challenging as it not only requires precisely aligning the composed query and target image in a common embedding space, but also simultaneously extracting related information from the reference image and modification text. In order to properly extract related information from the composed query, existing methods usually embed vision-language inputs using different feature encoders, e.g., CNN for images and LSTM/BERT for text, and then employ a complicated manually-designed composition module for learning the joint image-text representation. However, the architecture discrepancy in feature encoders would restrict the vision-language plenitudinous interaction. Meanwhile, certain complicated composition designs might significantly hamper the generalization ability of the model. To tackle these problems, we propose a new framework termed ComqueryFormer, which effectively processes the composed query with the Transformer for this task. Specifically, to eliminate the architecture discrepancy, we leverage a unified transformer-based architecture to homogeneously encode the vision-language inputs. Meanwhile, instead of the complicated composition module, the neat yet effective cross-modal transformer is adopted to hierarchically fuse the composed query at various vision scales. On the other hand, we introduce an efficient global-local alignment module to narrow the distance between the composed query and the target image. It not only considers the divergence in the global joint embedding space but also forces the model to focus on the local detail differences. Extensive experiments on three real-world datasets demonstrate the superiority of our ComqueryFormer. Our code can be found at:https://github.com/uestc-xyh/ComqueryFormer.

Cross-modal retrieval has recently attracted growing attention, which aims to match instances captured from different modalities. The performance of cross-modal retrieval methods heavily relies on the capability of metric learning to mine and weight the informative pairs. While various metric learning methods have been developed for unimodal retrieval tasks, the cross-modal retrieval tasks, however, have not been explored to its fullest extent. In this paper, we develop a universal weighting metric learning framework for cross-modal retrieval, which can effectively sample informative pairs and assign proper weight values to them based on their similarity scores so that different pairs favor different penalty strength. Based on this framework, we introduce two types of polynomial loss for cross-modal retrieval, self-similarity polynomial loss and relative-similarity polynomial loss. The former provides a polynomial function to associate the weight values with self-similarity scores, and the latter defines a polynomial function to associate the weight values with relative-similarity scores. Both self and relative-similarity polynomial loss can be freely applied to off-the-shelf methods and further improve their retrieval performance. Extensive experiments on two image-text retrieval datasets, three video-text retrieval datasets and one fine-grained image retrieval dataset demonstrate that our proposed method can achieve a noticeable boost in retrieval performance.

Cross-modal matching has attracted growing attention due to the rapid emergence of the multimedia data on the web and social applications. Recently, many re-weighting methods have been proposed for accelerating model training by designing a mapping function from similarity scores to weights. However, these re-weighting methods are difficult to be universally applied in practice since manually pre-set weighting functions inevitably involve hyper-parameters. In this paper, we propose a Meta Self-Paced Network (Meta-SPN) that automatically learns a weighting scheme from data for cross-modal matching. Specifically, a meta self-paced network composed of a fully connected neural network is designed to fit the weight function, which takes the similarity score of the sample pairs as input and outputs the corresponding weight value. Our meta self-paced network considers not only the self-similarity scores, but also their potential interactions (e.g., relative-similarity) when learning the weights. Motivated by the success of meta-learning, we use the validation set to update the meta self-paced network during the training of the matching network. Experiments on two image-text matching benchmarks and two video-text matching benchmarks demonstrate the generalization and effectiveness of our method.

The goal of Zero-shot Learning (ZSL) is to recognize categories that are not seen during the training process. The traditional method is to learn an embedding space and map visual features and semantic features to this common space. However, this method inevitably encounters the bias problem, i.e., unseen instances are often incorrectly recognized as the seen classes. Some attempts are made by proposing another paradigm, which uses generative models to hallucinate the features of unseen samples. However, the generative models often suffer from instability issues, making it impractical for them to generate fine-grained features of unseen samples, thus resulting in very limited improvement. To resolve this, a Semantic Enhanced Cross-modal GAN (SECM GAN) is proposed by imposing the cross-modal association for improving the semantic and discriminative property of the generated features. Specifically, we first train a cross-modal embedding model called Semantic Enhanced Cross-modal Model (SECM), which is constrained by discrimination and semantics. Then we train our generative model based on Generative Adversarial Network (GAN) called SECM GAN, in which the generator generates cross-modal features, and the discriminator distinguishes true cross-modal features from generated cross-modal features. We deploy SECM as a weak constraint of GAN, which makes reliance on GAN get reduced. We evaluate extensive experiments on three widely used ZSL datasets to demonstrate the superiority of our framework.

Cross-modal matching has been a highlighted research topic in both vision and language areas. Learning appropriate mining strategy to sample and weight informative pairs is crucial for the cross-modal matching performance. However, most existing metric learning methods are developed for unimodal matching, which is unsuitable for cross-modal matching on multimodal data with heterogeneous features. To address this problem, we propose a simple and interpretable universal weighting framework for cross-modal matching, which provides a tool to analyze the interpretability of various loss functions. Furthermore, we introduce a new polynomial loss under the universal weighting framework, which defines a weight function for the positive and negative informative pairs respectively. Experimental results on two image-text matching benchmarks and two video-text matching benchmarks validate the efficacy of the proposed method.

Zero-shot learning has been a highlighted research topic in both vision and language areas. Recently, generative methods have emerged as a new trend of zero-shot learning, which synthesizes unseen categories samples via generative models. However, the lack of fine-grained information in the synthesized samples makes it difficult to improve classification accuracy. It is also time-consuming and inefficient to synthesize samples and using them to train classifiers. To address such issues, we propose a novel Graph-based Variational Auto-Encoder for zero-shot learning. Specifically, we adopt knowledge graph to model the explicit inter-class relationships, and design a full graph convolution auto-encoder framework to generate the classifier from the distribution of the class-level semantic features on individual nodes. The encoder learns the latent representations of individual nodes, and the decoder generates the classifiers from latent representations of individual nodes. In contrast to synthesize samples, our proposed method directly generates classifiers from the distribution of the class-level semantic features for both seen and unseen categories, which is more straightforward, accurate and computationally efficient. We conduct extensive experiments and evaluate our method on the widely used large-scale ImageNet-21K dataset. Experimental results validate the efficacy of the proposed approach.

Scene graph generation intends to build graph-based representation from images, where nodes and edges respectively represent objects and relationships between them. However, scene graph generation today is heavily limited by imbalanced class prediction. Specifically, most of existing work achieves satisfying performance on simple and frequent relation classes (e.g. on), yet leaving poor performance with fine-grained and infrequent ones (e.g. walk on, stand on). To tackle this problem, in this paper, we redesign the framework as two branches, representation learning branch and classifier learning branch, for a more balanced scene graph generator. Furthermore, for representation learning branch, we propose Cross-modal Attention Coordinator (CAC) to gather consistent features from multi-modal using dynamic attention. For classifier learning branch, we first transfer relation classes' knowledge from large scale corpus, then we leverage Multi-Relationship classifier via Graph Attention neTworks (MR-GAT) to bridge the gap between frequent relations and infrequent ones. The comprehensive experimental results on VG200, a challenge dataset, indicate the competitiveness and the significant superiority of our proposed approach.

Most existing Zero-Shot Learning (ZSL) approaches adopt the semantic space as a bridge to classify unseen categories. However, it is difficult to transfer knowledge from seen categories to unseen categories through semantic space, since the correlations among categories are uncertain and ambiguous in the semantic space. In this paper, we formulated zero-shot learning as a classifier weight regression problem. Specifically, we propose a novel Residual Graph Convolution Network (ResGCN) which takes word embeddings and knowledge graph as inputs and outputs a visual classifier for each category. ResGCN can effectively alleviate the problem of over-smoothing and over-fitting. During the test, an unseen image can be classified by ranking the inner product of its visual feature and predictive visual classifiers. Moreover, we provide a new method to build a better knowledge graph. Our approach not only further enhances the correlations among categories, but also makes it easy to add new categories to the knowledge graph. Experiments conducted on the large-scale ImageNet 2011 21K dataset demonstrate that our method significantly outperforms existing state-of-the-art approaches.