Recent Multi-modal Large Language Models (MLLMs) have been challenged by the computational overhead resulting from massive video frames, often alleviated through compression strategies. However, the visual content is not equally contributed to user instructions, existing strategies (e.g., average pool) inevitably lead to the loss of potentially useful information. To tackle this, we propose the Hybridlevel Instruction Injection Strategy for Conditional Token Compression in MLLMs (HICom), utilizing the instruction as a condition to guide the compression from both local and global levels. This encourages the compression to retain the maximum amount of user-focused information while reducing visual tokens to minimize computational burden. Specifically, the instruction condition is injected into the grouped visual tokens at the local level and the learnable tokens at the global level, and we conduct the attention mechanism to complete the conditional compression. From the hybrid-level compression, the instruction-relevant visual parts are highlighted while the temporal-spatial structure is also preserved for easier understanding of LLMs. To further unleash the potential of HICom, we introduce a new conditional pre-training stage with our proposed dataset HICom-248K. Experiments show that our HICom can obtain distinguished video understanding ability with fewer tokens, increasing the performance by 2.43% average on three multiple-choice QA benchmarks and saving 78.8% tokens compared with the SOTA method. The code is available at https://github.com/lntzm/HICom.

Visual captioning benchmarks have become outdated with the emergence of modern multimodal large language models (MLLMs), as the brief ground-truth sentences and traditional metrics fail to assess detailed captions effectively. While recent benchmarks attempt to address this by focusing on keyword extraction or object-centric evaluation, they remain limited to vague-view or object-view analyses and incomplete visual element coverage. In this paper, we introduce CAPability, a comprehensive multi-view benchmark for evaluating visual captioning across 12 dimensions spanning six critical views. We curate nearly 11K human-annotated images and videos with visual element annotations to evaluate the generated captions. CAPability stably assesses both the correctness and thoroughness of captions with \textit{precision} and \textit{hit} metrics. By converting annotations to QA pairs, we further introduce a heuristic metric, \textit{know but cannot tell} ($K\bar{T}$), indicating a significant performance gap between QA and caption capabilities. Our work provides a holistic analysis of MLLMs' captioning abilities, as we identify their strengths and weaknesses across various dimensions, guiding future research to enhance specific aspects of their capabilities.

Generalist models have achieved remarkable success in both language and vision-language tasks, showcasing the potential of unified modeling. However, effectively integrating fine-grained perception tasks like detection and segmentation into these models remains a significant challenge. This is primarily because these tasks often rely heavily on task-specific designs and architectures that can complicate the modeling process. To address this challenge, we present UFO, a framework that unifies fine-grained visual perception tasks through an open-ended language interface. By transforming all perception targets into the language space, UFO unifies object-level detection, pixel-level segmentation, and image-level vision-language tasks into a single model. Additionally, we introduce a novel embedding retrieval approach that relies solely on the language interface to support segmentation tasks. Our framework bridges the gap between fine-grained perception and vision-language tasks, significantly simplifying architectural design and training strategies while achieving comparable or superior performance to methods with intricate task-specific designs. After multi-task training on five standard visual perception datasets, UFO outperforms the previous state-of-the-art generalist models by 12.3 mAP on COCO instance segmentation and 3.3 mIoU on ADE20K semantic segmentation. Furthermore, our method seamlessly integrates with existing MLLMs, effectively combining fine-grained perception capabilities with their advanced language abilities, thereby achieving superior performance on the challenging reasoning segmentation. Code and models are available at https://github.com/nnnth/UFO.

Zero-shot learning (ZSL) focuses on recognizing unseen categories by aligning visual features with semantic information. Recent advancements have shown that aligning each attribute with its corresponding visual region significantly improves zero-shot learning performance. However, the crude semantic proxies used in these methods fail to capture the varied appearances of each attribute, and are also easily confused by the presence of semantically redundant backgrounds, leading to suboptimal alignment. To combat these issues, we introduce a novel Alignment-Enhanced Network (AENet), designed to denoise the visual features and dynamically perceive semantic information, thus enhancing visual-semantic alignment. Our approach comprises two key innovations. (1) A visual denoising encoder, employing a class-agnostic mask to filter out semantically redundant visual information, thus producing refined visual features adaptable to unseen classes. (2) A dynamic semantic generator that crafts content-aware semantic proxies adaptively, steered by visual features, enabling AENet to discriminate fine-grained variations in visual contents. Additionally, we integrate a cross-fusion module to ensure comprehensive interaction between the denoised visual features and the generated dynamic semantic proxies, further facilitating visual-semantic alignment. Through extensive experiments across three datasets, the proposed method demonstrates that it narrows down the visual-semantic gap and sets a new benchmark in this setting.

Video Moment Retrieval (VMR) aims to retrieve temporal segments in untrimmed videos corresponding to a given language query by constructing cross-modal alignment strategies. However, these existing strategies are often sub-optimal since they ignore the modality imbalance problem, i.e., the semantic richness inherent in videos far exceeds that of a given limited-length sentence. Therefore, in pursuit of better alignment, a natural idea is enhancing the video modality to filter out query-irrelevant semantics, and enhancing the text modality to capture more segment-relevant knowledge. In this paper, we introduce Modal-Enhanced Semantic Modeling (MESM), a novel framework for more balanced alignment through enhancing features at two levels. First, we enhance the video modality at the frame-word level through word reconstruction. This strategy emphasizes the portions associated with query words in frame-level features while suppressing irrelevant parts. Therefore, the enhanced video contains less redundant semantics and is more balanced with the textual modality. Second, we enhance the textual modality at the segment-sentence level by learning complementary knowledge from context sentences and ground-truth segments. With the knowledge added to the query, the textual modality thus maintains more meaningful semantics and is more balanced with the video modality. By implementing two levels of MESM, the semantic information from both modalities is more balanced to align, thereby bridging the modality gap. Experiments on three widely used benchmarks, including the out-of-distribution settings, show that the proposed framework achieves a new start-of-the-art performance with notable generalization ability (e.g., 4.42% and 7.69% average gains of [email protected] on Charades-STA and Charades-CG). The code will be available at https://github.com/lntzm/MESM.

Generalized Zero-Shot Learning (GZSL) aims to recognize both seen and unseen categories by establishing visual and semantic relations. Recently, generation-based methods that focus on synthesizing fictitious visual features from corresponding attributes have gained significant attention. However, these generated features often lack discriminative capabilities due to inadequate training of the generative model. To address this issue, we propose a novel Discriminative Enhanced Network (DENet) to harness the potential of the generative model by adapting the training features and imposing constraints on the generated features. Our approach incorporates three pivotal modules: (1) Before the generative network training, we implement a Pre-Tuning Module (PTM) to eliminate irrelevant background noise in the raw features extracted from a fixed CNN backbone. Therefore, PTM can provide tuned training features without redundant noise for generative model. (2) During the generative network training, we propose an Asymmetry Cross-authenticity Contrastive (AC2) loss to group visual features of the same category while repel features from different categories by optimizing a large number of sample pairs. Additionally, we incorporate intra-class and relation-specific inter-class boundaries within the AC2 loss to enrich sample diversity and preserve valid semantic information. (3) Also within the generative network training, a Dual-semantic Alignment Module (DAM) is designed to align visual features with both attributes and label embeddings, enabling the model to learn attribute-related information and discriminative extended semantics. Experiments on four standard benchmarks demonstrate that our approach learns more discriminative features and surpasses the existing methods.

Conventional detectors suffer from performance degradation when dealing with long-tailed data due to a classification bias towards the majority head categories. In this article, we contend that the learning bias originates from two factors: 1) the unequal competition arising from the imbalanced distribution of foreground categories, and 2) the lack of sample diversity in tail categories. To tackle these issues, we introduce a unified framework calledBAlancedCLassification (BACL), which enables adaptive rectification of inequalities caused by disparities in category distribution and dynamic intensification of sample diversities in a synchronized manner. Specifically, a novel foreground classification balance loss (FCBL) is developed to ameliorate the domination of head categories and shift attention to difficult-to-differentiate categories by introducing pairwise class-aware margins and auto-adjusted weight terms, respectively. This loss prevents the over-suppression of tail categories in the context of unequal competition. Moreover, we propose a dynamic feature hallucination module (FHM), which enhances the representation of tail categories in the feature space by synthesizing hallucinated samples to introduce additional data variances. In this divide-and-conquer approach, BACL sets a new state-of-the-art on the challenging LVIS benchmark with a decoupled training pipeline, surpassing vanilla Faster R-CNN with ResNet-50-FPN by 5.8% AP and 16.1% AP for overall and tail categories. Extensive experiments demonstrate that BACL consistently achieves performance improvements across various datasets with different backbones and architectures.

Hashing has been studied extensively for large-scale image retrieval due to its efficient computation and storage. Deep hashing methods typically train models with category-balanced data and suffer from a serious performance deterioration when dealing with long-tailed training samples. Recently, several long-tailed hashing methods focus on this newly emerging field for practical purpose. However, existing methods still face challenges that fixed category centers with limited semantic information cannot effectively improve the discriminative ability of tail-category hash codes. To tackle the issue, we propose a novel method called Semantic-enhanced Proxy-guided Hashing in this paper. We leverage two sets of learnable category proxies in the feature space and the Hamming space respectively, which can describe category semantics by getting updated continuously along with the whole model via back-propagation. Based on this, we introduce the Mahalanobis distance metric to characterize relationships accurately and enhance the semantic representation of both proxies and samples concurrently, improving the hash learning process. Moreover, we capture the multilateral correlations between proxies and samples in the feature space and extend a hypergraph neural network to transfer semantic knowledge from proxies to samples in the Hamming space. Extensive experiments show that our method achieves the state-of-the-art performance and surpasses existing methods by 1.47%–7.56% MAP on long-tailed benchmarks, demonstrating the superiority of learnable category proxies and the effectiveness of our proposed learning algorithm for long-tailed hashing.

The performance of text-video retrieval has been significantly improved by vision-language cross-modal learning schemes. The typical solution is to directly align the global video-level and sentence-level features during learning, which would ignore the intrinsic video-text relations, i.e., a text description only corresponds to a spatio-temporal part of videos. Hence, the matching process should consider both fine-grained spatial content and various temporal semantic events. To this end, we propose a text-video learning framework with progressive spatio-temporal prototype matching. Specifically, the matching process is decomposed into two complementary phases: object-phrase prototype matching and event-sentence prototype matching. In the object-phrase prototype matching phase, the spatial prototype generation mechanism predicts key patches or words, which are aggregated into object or phrase prototypes. Importantly, optimizing the local alignment between object-phrase prototypes helps the model perceive spatial details. In the event-sentence prototype matching phase, we design a temporal prototype generation mechanism to associate intra-frame objects and interact inter-frame temporal relations. Such progressively generated event prototypes can reveal semantic diversity in videos for dynamic matching. Validated by comprehensive experiments, our method consistently outperforms the state-of-the-art methods on four video retrieval benchmark.1

Deep hashing has been extensively explored for image retrieval due to fast computation and efficient storage. Since conventional deep hashing methods are not suitable for the common scenario in real life that data exhibits a long-tailed distribution, several long-tailed hashing methods have been proposed recently. However, existing long-tail hashing methods seek to utilize fixed class centroids and cannot fully develop the discriminative ability of hash codes for tail-class samples. Specifically, fixed class centroids cannot characterize authentic semantics of tail classes or provide effective semantic information for hash codes learning under the long-tailed setting. To this end, we propose a novel Dual Dynamic Proxy Hashing Network (DDPHN) with two sets of learnable dynamic proxies, i.e. hash proxies and feature proxies, to improve the discrimination of hash codes for tail-class samples. Compared with fixed class centroids, learnable proxies can be optimized constantly via the proxy learning loss and depict accurate class semantics despite the scarcity of tail-class samples. Apart from low-dimensional binary hash proxies, we introduce high-dimensional continuous feature proxies that can describe semantic relationships more precisely, contributing to hash codes learning as well. To further leverage semantic information carried by proxies, we build a hypergraph by exploring neighborhood relationships in the feature space and then introduce a hypergraph neural network to transfer knowledge from proxies to samples in the Hamming space. Extensive experiments show the superiority of our learnable dynamic proxies and demonstrate that our method outperforms numerous deep hashing models and recent state-of-the-art long-tailed hashing methods.

Zero-Shot Learning (ZSL) aims to recognize images from seen and unseen classes by aligning visual and semantic knowledge (e.g., attribute descriptions). However, the fine-grained attributes in the RGB domain can be easily affected by background noise (e.g., the grey bird tail blending with the ground), making it difficult to effectively distinguish them. Analyzing the features in the frequency domain assists in better distinguishing the attributes since their patterns remain consistent across different images, unlike noise which may be more variable. Nevertheless, existing ZSL methods typically learn visual features directly from the RGB domain, which can impede the recognition of certain attributes. To overcome this limitation, we propose a novel ZSL method named Frequency-based Phase Augmentation (FPA) network, which learns an effective representation of the attributes in the frequency domain. Specifically, we introduce a Hybrid Phase Augmentation (HPA) module to transform visual features into the frequency domain and augment the phase component for better retention of semantic information of the attributes. The use of phase-augmented features enables FPA to capture more semantic knowledge that can be challenging to distinguish in the RGB domain, suppress noise, and highlight significant attributes. Our extensive experiments show that FPA achieves state-of-the-art performance across four standard datasets.

Video moment retrieval pursues an efficient and generalized solution to identify the specific temporal segments within an untrimmed video that correspond to a given language description. To achieve this goal, we provide a generative diffusion-based framework called MomentDiff, which simulates a typical human retrieval process from random browsing to gradual localization. Specifically, we first diffuse the real span to random noise, and learn to denoise the random noise to the original span with the guidance of similarity between text and video. This allows the model to learn a mapping from arbitrary random locations to real moments, enabling the ability to locate segments from random initialization. Once trained, MomentDiff could sample random temporal segments as initial guesses and iteratively refine them to generate an accurate temporal boundary. Different from discriminative works (e.g., based on learnable proposals or queries), MomentDiff with random initialized spans could resist the temporal location biases from datasets. To evaluate the influence of the temporal location biases, we propose two ``anti-bias'' datasets with location distribution shifts, named Charades-STA-Len and Charades-STA-Mom. The experimental results demonstrate that our efficient framework consistently outperforms state-of-the-art methods on three public benchmarks, and exhibits better generalization and robustness on the proposed anti-bias datasets. The code, model, and anti-bias evaluation datasets will be released publicly.

Deep metric learning methods generally concentrate on designing distance-based losses to learn sample embeddings, which tacitly presuppose the neighborhood structure around each sample (e.g., hypersphere for Euclidean distance). However, this supposition is overly optimistic: 1) visual data is often located on low-dimensional manifolds curved in high-dimensional space, and all regions of the manifold may hardly share the same local structures in the input space; 2) it is unlikely that the local structure in the output embedding space is as homogeneous as assumed due to the non-linearity of neural networks. Hence, simply characterizing sample embeddings while ignoring the respective neighborhood structures leads to limitations. To address this problem, this paper presents a Neighborhood-Adaptive Multi-cluster Ranking (NAMR) framework by leveraging the heterogeneity of local structures. Specifically, considering that indexing algorithms are usually required in large-scale retrieval, NAMR characterizes an image from two kinds of embeddings (i.e., sample embedding and structure embedding). The sample embedding can be trained using any distance-based loss, while the structure embedding representing the neighborhood structure can be jointly learned with the sample embedding in a self-supervised multi-cluster ranking manner. In this way, existing indexing algorithms can seamlessly support large-scale retrieval employing NAMR embeddings without any modifications. We evaluate the proposed model on five standard benchmarks, consistently and explicitly improving four baselines (especially the simplest triplet loss) and achieving state-of-the-art performance.

Most metric learning techniques typically focus on sample embedding learning, while implicitly assume a homogeneous local neighborhood around each sample, based on the metrics used in training ( e.g., hypersphere for Euclidean distance or unit hyperspherical crown for cosine distance). As real-world data often lies on a low-dimensional manifold curved in a high-dimensional space, it is unlikely that everywhere of the manifold shares the same local structures in the input space. Besides, considering the non-linearity of neural networks, the local structure in the output embedding space may not be homogeneous as assumed. Therefore, representing each sample simply with its embedding while ignoring its individual neighborhood structure would have limitations in Embedding-Based Retrieval (EBR). By exploiting the heterogeneity of local structures in the embedding space, we propose a Neighborhood-Adaptive Structure Augmented metric learning framework (NASA), where the neighborhood structure is realized as a structure embedding, and learned along with the sample embedding in a self-supervised manner. In this way, without any modifications, most indexing techniques can be used to support large-scale EBR with NASA embeddings. Experiments on six standard benchmarks with two kinds of embeddings, i.e., binary embeddings and real-valued embeddings, show that our method significantly improves and outperforms the state-of-the-art methods.

Zero-Shot Learning (ZSL) aims to recognize unseen classes by transferring knowledge from seen classes. Recent methods focus on learning a common semantic space to align visual and attribute information. However, they always over-relied on provided attributes and ignored the category discriminative information that contributes to accurate unseen class recognition, resulting in weak transferability. To this end, we propose a novel Dual Part Discovery Network (DPDN) that considers both attribute and category discriminative information by discovering attribute-guided parts and category-guided parts simultaneously to improve knowledge transfer. Specifically, for attribute-guided parts discovery, DPDN can localize the regions with specific attribute information and significantly bridge the gap between visual and semantic information guided by the given attributes. For category-guided parts discovery, the local parts are explored to discover other important regions that bring latent crucial details ignored by attributes, with the guidance of adaptive category prototypes. To better mine the transferable knowledge, we impose class correlations constraints to regularize the category prototypes. Finally, attribute- and category-guided parts complement each other and provide adequate discriminative subtle information for more accurate unseen class recognition. Extensive experimental results demonstrate that DPDN can discover discriminative parts and outperform state-of-the-art methods on three standard benchmarks.

To reduce the extreme label dependence of supervised product quantization methods, the semi-supervised paradigm usually employs massive unlabeled data to assist in regularizing deep networks, thereby improving model performance. However, the existing method focuses on the overall distribution consistency between unlabeled data and class prototypes, while ignoring subtle individual variances between unlabeled instances. Therefore, the local neighborhood structure is not fully explored, which will cause the model to easily overfit in the training set. In this paper, we introduce a new Fourier perspective to alleviate this issue by exploring the semantic relations between unlabeled instances in a self-supervised manner. Specifically, based on Fourier Transform, we first design a Phase Mixing (PM) strategy, which can manipulate the mixing area and values of the phase component between two images to control the proportion of semantic information. In this way, we can construct multi-level similarity neighbors naturally for unlabeled data. Then, a ranking quantization loss is formulated to perceive multi-level semantic variances in neighbor instances, which improves the robustness and generalization of the model. Extensive experiments in three different semi-supervised settings show that our method outperforms existing state-of-the-art methods by averaged 3.95% improvement on four datasets.

Recently, the online retrieval task has been receiving widespread attention, which is closely related to many real-world applications. However, existing online retrieval methods based on hashing suffer from two main problems: a) the models tend to be biased towards the current streaming data due to unavailable history streaming data; b) when new streaming data comes in and the hashing functions have been updated, all history binary codes should be recomputed, which takes much computation burden. To address the above two issues, we propose a novel Online Residual Quantization (ORQ) method that can achieve efficient streaming data quantization via the small-scale residual quantization codebooks. For the first problem, we design a residual quantization module by learning multiple residual codebooks to quantize the float streaming data, which effectively reduces the quantization error and enables the binary codes to be easily reconstructed back to original float data. Then, with the reconstructed history data, a balanced affinity matrix is developed to model the semantic relationship,e.g.,similarity and difference, between the history and current data distributions, which can prevent the model from being biased towards the current data distribution. For the second problem, when inputting current streaming data, only the residual codebooks should be updated, instead of the whole history binary codes in hashing-based methods, which significantly reduces the computation burden. Comprehensive experiments on six benchmarks demonstrate that ORQ yields significant improvements (i.e.,1.2%$\sim$4.9% in average mAP) compared to the state-of-the-art methods.