Vision-language pre-training (VLP) on large-scale datasets has shown premier performance on various downstream tasks. In contrast to plenty of available benchmarks with English corpus, large-scale pre-training datasets and downstream datasets with Chinese corpus remain largely unexplored. In this work, we build a large-scale high-quality Chinese Cross-Modal Benchmark named CCMB for the research community, which contains the currently largest public pre-training dataset Zero and five human-annotated fine-tuning datasets for downstream tasks. Zero contains 250 million images paired with 750 million text descriptions, plus two of the five fine-tuning datasets are also currently the largest ones for Chinese cross-modal downstream tasks. Along with the CCMB, we also develop a VLP framework named R2D2, applying a pre-Ranking + Ranking strategy to learn powerful vision-language representations and a two-way distillation method (i.e., target-guided Distillation and feature-guided Distillation) to further enhance the learning capability. With the Zero and the R2D2 VLP framework, we achieve state-of-the-art performance on twelve downstream datasets from five broad categories of tasks including image-text retrieval, image-text matching, image caption, text-to-image generation, and zero-shot image classification. The datasets, models, and codes are available at https://github.com/yuxie11/R2D2
论文检索
输入标题、作者或关键词，从 1,014 篇学术成果中精准定位
Focusing on Flexible Masks: A Novel Framework for Panoptic Scene Graph Generation with Relation Constraints
PDF ↗Panoptic Scene Graph Generation (PSG) presents pixel-wise instance detection and localization, leading to comprehensive and precise scene graphs. Current methods employ conventional Scene Graph Generation (SGG) frameworks to solve the PSG problem, neglecting the fundamental differences between bounding boxes and masks, i.e., bounding boxes are allowed overlap but masks are not. Since segmentation from the panoptic head has deviations, non-overlapping masks may not afford complete instance information. Subsequently, in the training phase, incomplete segmented instances may not be well-aligned to annotated ones, causing mismatched relations and insufficient training. During the inference phase, incomplete segmentation leads to incomplete scene graph prediction. To alleviate these problems, we construct a novel two-stage framework for the PSG problem. In the training phase, we design a proposal matching strategy, which replaces deterministic segmentation results with proposals extracted from the off-the-shelf panoptic head for label alignment, thereby ensuring the all-matching of training samples. In the inference phase, we present an innovative concept of employing relation predictions to constrain segmentation and design a relation-constrained segmentation algorithm. By reconstructing the process of generating segmentation results from proposals using predicted relation results, the algorithm recovers more valid instances and predicts more complete scene graphs. The experimental results show overall superiority, effectiveness, and robustness against adversarial attacks.
Semi-supervised semantic segmentation (SSS) is an important task that utilizes both labeled and unlabeled data to reduce expenses on labeling training examples. However, the effectiveness of SSS algorithms is limited by the difficulty of fully exploiting the potential of unlabeled data. To address this, we propose a dual-level Siamese structure network (DSSN) for pixel-wise contrastive learning. By aligning positive pairs with a pixel-wise contrastive loss using strong augmented views in both low-level image space and high-level feature space, the proposed DSSN is designed to maximize the utilization of available unlabeled data. Additionally, we introduce a novel class-aware pseudo-label selection strategy for weak-to-strong supervision, which addresses the limitations of most existing methods that do not perform selection or apply a predefined threshold for all classes. Specifically, our strategy selects the top high-confidence prediction of the weak view for each class to generate pseudo labels that supervise the strong augmented views. This strategy is capable of taking into account the class imbalance and improving the performance of long-tailed classes. Our proposed method achieves state-of-the-art results on two datasets, PASCAL VOC 2012 and Cityscapes, outperforming other SSS algorithms by a significant margin. The source code is available at https://github.com/kunzhan/DSSN.
This paper addresses the challenging task of language-driven moment retrieval. Previous methods are typically trained to localize the target moment corresponding to a single sentence query in a complicated video. However, this specific moment generally delivers richer contents than the query, i.e., the semantics of one query may miss certain object details or actions in the complex foreground-background visual contents. Such information imbalance between two modalities makes it difficult to finely align their representations. To this end, instead of training with a single query, we propose to utilize the diversity and complementarity among different queries corresponding to the same video moment for enriching the textual semantics. Specifically, we develop a Teacher-Student Moment Retrieval (TSMR) framework to fill this cross-modal information gap. A teacher model is trained to not only encode a certain query but also capture extra complementary queries to aggregate contextual semantics for obtaining more comprehensive moment-related query representations. Since the additional queries are inaccessible during inference, we further introduce an adaptive knowledge distillation mechanism to train a student model with a single query input by selectively absorbing the knowledge from the teacher model. In this manner, the student model is more robust to the cross-modal information gap during the moment retrieval guided by a single query. Experimental results on two benchmarks demonstrate the effectiveness of our proposed method.
Despite the broad application of Machine Learning models as a Service (MLaaS), they are vulnerable to model stealing attacks. These attacks can replicate the model functionality by using the black-box query process without any prior knowledge of the target victim model. Existing stealing defenses add deceptive perturbations to the victim's posterior probabilities to mislead the attackers. However, these defenses are now suffering problems of high inference computational overheads and unfavorable trade-offs between benign accuracy and stealing robustness, which challenges the feasibility of deployed models in practice. To address the problems, this paper proposes Isolation and Induction (InI), a novel and effective training framework for model stealing defenses. Instead of deploying auxiliary defense modules that introduce redundant inference time, InI directly trains a defensive model by isolating the adversary's training gradient from the expected gradient, which can effectively reduce the inference computational cost. In contrast to adding perturbations over model predictions that harm the benign accuracy, we train models to produce uninformative outputs against stealing queries, which can induce the adversary to extract little useful knowledge from victim models with minimal impact on the benign performance. Extensive experiments on several visual classification datasets (e.g., MNIST and CIFAR10) demonstrate the superior robustness (up to 48% reduction on stealing accuracy) and speed (up to 25.4× faster) of our InI over other state-of-the-art methods. Our codes can be found in https://github.com/DIG-Beihang/InI-Model-Stealing-Defense.
Point cloud generation is an important research topic in 3D computer vision, which can provide high-quality datasets for various downstream tasks. However, efficiently capturing the geometry of point clouds remains a challenging problem due to their irregularities. In this paper, we propose a novel transformer-based 3D point cloud generation network to generate realistic point clouds. Specifically, we first develop a transformer-based interpolation module that utilizes k-nearest neighbors at different scales to learn global and local information about point clouds in the feature space. Based on geometric information, we interpolate new point features to upsample the point cloud features. Then, the upsampled features are used to generate a coarse point cloud with spatial coordinate information. We construct a transformer-based refinement module to enhance the upsampled features in feature space with geometric information in coordinate space. Finally, we use a multi-layer perceptron on the upsampled features to generate the final point cloud. Extensive experiments on ShapeNet and ModelNet demonstrate the effectiveness of our proposed method.
Text-based person retrieval (TPR) is a challenging task that involves retrieving a specific individual based on a textual description. Despite considerable efforts to bridge the gap between vision and language, the significant differences between these modalities continue to pose a challenge. Previous methods have attempted to align text and image samples in a modal-shared space, but they face uncertainties in optimization directions due to the movable features of both modalities and the failure to account for one-to-many relationships of image-text pairs in TPR datasets. To address this issue, we propose an effective bi-directional one-to-many embedding paradigm that offers a clear optimization direction for each sample, thus mitigating the optimization problem. Additionally, this embedding scheme generates multiple features for each sample without introducing trainable parameters, making it easier to align with several positive samples. Based on this paradigm, we propose a novel Bi-directional one-to-many Embedding Alignment (Beat) model to address the TPR task. Our experimental results demonstrate that the proposed Beat model achieves state-of-the-art performance on three popular TPR datasets, including CUHK-PEDES (65.61 R@1), ICFG-PEDES (58.25 R@1), and RSTPReID (48.10 R@1). Furthermore, additional experiments on MS-COCO, CUB, and Flowers datasets further demonstrate the potential of Beat to be applied to other image-text retrieval tasks.
Video restoration networks aim to restore high-quality frame sequences from degraded ones. However, traditional video restoration methods heavily rely on temporal modeling operators or optical flow estimation, which limits their versatility. The aim of this work is to present a novel approach for video restoration that eliminates inefficient temporal modeling operators and pixel-level feature alignment in the network architecture. The proposed method, Sequential Affinity Learning Network (SALN), is designed based on an affinity mechanism that establishes direct correspondences between the Query frame, degraded sequence, and restored frames in latent space. This unique perspective allows for more accurate and effective restoration of video content without relying on temporal modeling operators or optical flow estimation techniques. Moreover, we enhanced the design of the channel-wise self-attention block to improve the decoder's performance for video restoration. Our method outperformed previous state-of-the-art methods by a significant margin in several classic video tasks, including video deraining, video dehazing, and video waterdrop removal, demonstrating excellent efficiency. As a novel network that differs significantly from previous video restoration methods, SALN aims to provide innovative ideas and directions for video restoration. Our contributions include proposing a novel affinity-based approach for video restoration, enhancing the design of the channel-wise self-attention block, and achieving state-of-the-art performance on several classic video tasks.
Zero-Shot Learning (ZSL) aims to recognize unseen classes by generalizing the knowledge, i.e., visual and semantic relationships, obtained from seen classes, where image augmentation techniques are commonly applied to improve the generalization ability of a model. However, this approach can also cause adverse effects on ZSL since the conventional augmentation techniques that solely depend on single-label supervision is not able to maintain semantic information and result in the semantic distortion issue consequently. In other words, image argumentation may falsify the semantic (e.g., attribute) information of an image. To take the advantage of image augmentations while mitigating the semantic distortion issue, we propose a novel ZSL approach by Harnessing Adversarial Samples (HAS). HAS advances ZSL through adversarial training which takes into account three crucial aspects: (1) robust generation by enforcing augmentations to be similar to negative classes, while maintaining correct labels, (2) reliable generation by introducing a latent space constraint to avert significant deviations from the original data manifold, and (3) diverse generation by incorporating attribute-based perturbation by adjusting images according to each semantic attribute's localization. Through comprehensive experiments on three prominent zero-shot benchmark datasets, we demonstrate the effectiveness of our adversarial samples approach in both ZSL and Generalized Zero-Shot Learning (GZSL) scenarios. Our source code is available at https://github.com/uqzhichen/HASZSL.
Existing superpixel segmentation algorithms mainly focus on natural image with high-quality, while neglecting the inevitable environment constraint in complex scenes. In this paper, we propose an end-to-end frequency domain guided superpixel segmentation network (FSNet) to generate superpixels with sharp boundary adherence for complex scenes by fusing the deep features in spatial and frequency domains. To utilize the frequency domain information of the image, an improved frequency information extractor (IFIE) is proposed to extract the frequency domain information with sharp boundary features. Moreover, considering the over-sharp feature may damage the semantic information of superpixel, we further design a dense hybrid atrous convolution (DHAC) block to preserve semantic information via capturing wider and deeper semantic information in spatial domain. Finally, the extracted deep features in spatial and frequency domains will be fused to generate semantic perceptual superpixels with sharp boundary adherence. Extensive experiments on multiple challenging datasets with complex boundaries demonstrate that our method achieves the state-of-the-art performance both quantitatively and qualitatively, and we further verify the superiority of the proposed method when applied in salient object detection.
Nighttime image dehazing is a challenging task due to the presence of multiple types of adverse degrading effects including glow, haze, blur, noise, color distortion, and so on. However, most previous studies mainly focus on daytime image dehazing or partial degradations presented in nighttime hazy scenes, which may lead to unsatisfactory restoration results. In this paper, we propose an end-to-end transformer-based framework for nighttime haze removal, called NightHazeFormer. Our proposed approach consists of two stages: supervised pre-training and semi-supervised fine-tuning. During the pre-training stage, we introduce two powerful priors into the transformer decoder to generate the non-learnable prior queries, which guide the model to extract specific degradations. For the fine-tuning, we combine the generated pseudo ground truths with input real-world nighttime hazy images as paired images and feed into the synthetic domain to fine-tune the pre-trained model. This semi-supervised fine-tuning paradigm helps improve the generalization to real domain. In addition, we also propose a large-scale synthetic dataset called UNREAL-NH, to simulate the real-world nighttime haze scenarios comprehensively. Extensive experiments on several synthetic and real-world datasets demonstrate the superiority of our NightHazeFormer over state-of-the-art nighttime haze removal methods in terms of both visually and quantitatively.
Video moment retrieval (VMR) aims to search for a video segment that matches the search intent in a query sentence, which has received increasing attention in recent years, due to its practical values in various fields. Existing efforts devoted to this interesting yet challenging task typically encode the query sentence and video segments into unstructured global representations for cross-modal interaction and fusion, which may fail to accurately capture the search intent in complex queries with multi-granularity semantics. To fill the research gap, this paper presents a novel solution termed semantics-enriched video moment retrieval method (SVMR), which can effectively and explicitly model the hierarchical multi-granularity semantics of complex textual query. Specifically, we first explore cross-token relations to offer multiple granularity query representations with hierarchical semantic contexts of semantically associated tokens for fine-grained cross-modal interaction and fusion, which contributes to mining rich visual motion cues semantically related to different activities and entities in complex queries. Furthermore, to fully leverage fine-grained cross-modal cues for moment retrieval, we design a specific temporal boundary reasoning module by explicitly generating start and end time-aware filter kernels with visual cues to perceive the moment boundaries. Extensive experiments and analyses on three public benchmarks clearly demonstrate the advantage of our proposed SVMR over existing state-of-the-art approaches, especially in retrieving complex query-based video moments.
Bird's-Eye-View (BEV) based 3D visual perception, which formulates a unified space for multi-view representation, has received wide attention in autonomous driving due to its scalability for downstream tasks. However, view transform in transformer-based BEV methods is agnostic of 3D occlusion relationships, resulting in model degradation. To construct a higher-quality BEV space, this paper analyzes the mutual occlusion problems in the view transform process and proposes a new transformer-based method named OccluBEV. OccluBEV alleviates the occlusion issue via point cloud information distillation in both the image and BEV space. Specifically, in the image space, we perform depth estimation for each pixel and utilize it to guide image feature mapping. Further, since predicting depth directly from monocular image is ill-posed, ignoring stereo information such as multi-view and temporal cues, this paper introduces a voxel visibility segmentation task in 3D BEV space. The task explicitly predicts whether each voxel in the 3D BEV grid is occupied or not. In addition, to alleviate the overfitting problem in BEV feature learning under a single task, we design a multi-head learning framework which jointly models multiple strongly-correlated tasks in a unified BEV space. The effectiveness of the proposed method is fully validated on the nuScenes dataset, achieving a competetive NDS/mAP score of 57.5/47.9 on the nuScenes test leaderboard using ResNet101 backbone, which is superior to state-of-the-art camera-based solutions.
Considerable performance gains have been achieved for knowledge-based visual question answering due to the visual-language pre-training models with pre-training-then-fine-tuning paradigm. However, because the targets of the pre-training and fine-tuning stages are different, there is an evident barrier that prevents the cross-modal comprehension ability developed in the pre-training stage from fully endowing the fine-tuning task. To break this barrier, in this paper, we propose a novel hybrid prompting model for knowledge-based VQA, which inherits and incorporates the pre-training and fine-tuning tasks with a shared objective. Specifically, based on static declaration prompt, we construct a consistent goal with the fine-tuning via masked language modeling to inherit capabilities of pre-training task, while selecting the top-t relevant knowledge in a dense retrieval manner. Additionally, a dynamic knowledge prompt is learned from retrieved knowledge, which not only alleviates the length constraint on inputs for visual-language pre-trained models but also assists in providing answer features via fine-tuning. Combining and unifying the aims of the two stages could fully exploit the abilities of pre-training and fine-tuning to predict answer. We evaluate the proposed model on the OKVQA dataset, and the result shows that our model outperforms the state-of-the-art methods based on visual-language pre-training models with a noticeable performance gap and even exceeds the large-scale language model of GPT-3, which proves the benefits of the hybrid prompts and the advantages of unifying pre-training to fine-tuning.
Knowledge graph (KG) schema, which prescribes a high-level structure and semantics of a KG, is significantly helpful for KG completion and reasoning problems. Despite its usefulness, open-domain KGs do not practically have a unified and fixed schema. Existing approaches usually extract schema information using entity types from a KG where each entity e can be associated with a set of types {Te, by either heuristically taking one type for each entity or exhaustively combining the types of all entities in a fact (to get entity-typed tuples, (h_type, r, t_type) for example). However, these two approaches either overlook the role of multiple types of a single entity across different facts or introduce non-negligible noise as not all the type combinations actually support the fact, thus failing to capture the sophisticated schema information. Against this background, we study the problem of modeling hyper-relational schema, which is formulated as mixed hyper-relational tuples ({Th}, r, {Tt}, k, {Tv1},...) with two-fold hyper-relations: each type set T may contain multiple types and each schema tuple may contain multiple key-type set pairs (k, Tv). To address this problem, we propose HELIOS, a hyper-relational schema model designed to subtly learn from such hyper-relational schema tuples by capturing not only the correlation between multiple types of a single entity, but also the correlation between types of different entities and relations in a schema tuple. We evaluate HELIOS on three real-world KG datasets in different schema prediction tasks. Results show that HELIOS consistently outperforms state-of-the-art hyper-relational link prediction techniques by 20.0-29.7%, and is also much more robust than baselines in predicting types and relations across different positions in a hyper-relational schema tuple.
Induction Network: Audio-Visual Modality Gap-Bridging for Self-Supervised Sound Source Localization
PDF ↗Self-supervised sound source localization is usually challenged by the modality inconsistency. In recent studies, contrastive learning based strategies have shown promising to establish such a consistent correspondence between audio and sound sources in visual scenarios. Unfortunately, the insufficient attention to the heterogeneity influence in the different modality features still limits this scheme to be further improved, which also becomes the motivation of our work. In this study, an Induction Network is proposed to bridge the modality gap more effectively. By decoupling the gradients of visual and audio modalities, the discriminative visual representations of sound sources can be learned with the designed Induction Vector in a bootstrap manner, which also enables the audio modality to be aligned with the visual modality consistently. In addition to a visual weighted contrastive loss, an adaptive threshold selection strategy is introduced to enhance the robustness of the Induction Network. Substantial experiments conducted on SoundNet-Flickr and VGG-Sound Source datasets have demonstrated a superior performance compared to other state-of-the-art works in different challenging scenarios. The code is available at https://github.com/Tahy1/AVIN.
Audio-visual speaker diarization refers to the task of identifying "who spoke when" by using both audio and video data. Although previous fusion-based approaches have shown exceptional performance over audio-only methods, they have mainly focused on high-quality data and have not accounted for the impacts of acoustic noise or missing faces. To address these limitations, we propose a novel uncertainty-aware end-to-end audio-visual speaker diarization (UAV-SD) approach in this paper. Our approach leverages both framewise inter- and intra-modal confidence to achieve more effective and robust speaker diarization. By taking into account the uncertainty of the data, UAV-SD can achieve better diarization performance even in noisy or low-quality recordings. Additionally, our approach is compatible with multi-channel audio signals without the need to retrain the model, making it a more versatile solution. To evaluate the effectiveness of our approach, we conduct extensive experiments on the Multi-modal Information Based Speech Processing (MISP) 2022 Challenge datasets which consist of far-field audio and video data. The results show that UAV-SD is able to yield significant performance gains compared to baseline methods for both single and multi-channel data, demonstrating its effectiveness in real-world scenarios.
Cross-network node classification aims to leverage the abundant knowledge from a labeled source network to help classify the node in an unlabeled target network. However, existing methods assume that label sets are identical across domains, which is easily violated in practice. Hence, we attempt to integrate network embedding with universal domain adaptation, which transfers valuable knowledge across domains without assumption on the label sets, to assist in node classification. Nonetheless, the complex network relationships between nodes increase the difficulty of this universal domain adaptive node classification task. In this work, we propose a novel Universal Domain Adaptive Network Embedding (UDANE) framework, which learns transferable node representations across networks to succeed in such a task. Technically, we first adopt the cross-network node embedding component to model comprehensive node information of both networks. Then we employ the inter-domain adaptive alignment component to exploit and relate knowledge across domains, learning domain-invariant representation for knowledge transfer. In addition, the intra-domain contrastive alignment component is proposed to learn discriminative representations beneficial for classification by sufficiently utilizing unlabeled data in the target domain. Extensive experiments have been conducted on real-world datasets, demonstrating that the proposed UDANE model outperforms the state-of-the-art baselines by a large margin.
In recent years, open-vocabulary (OV) object detection has attracted increasing research attention. Unlike traditional detection, which only recognizes fixed-category objects, OV detection aims to detect objects in an open category set. Previous works often leverage vision-language (VL) training data (e.g., referring grounding data) to recognize OV objects. However, they only use pairs of nouns and individual objects in VL data, while these data usually contain much more information, such as scene graphs, which are also crucial for OV detection. In this paper, we propose a novel Scene-Graph-Based Discovery Network (SGDN) that exploits scene graph cues for OV detection. Firstly, a scene-graph-based decoder (SGDecoder) including sparse scene-graph-guided attention (SSGA) is presented. It captures scene graphs and leverages them to discover OV objects. Secondly, we propose scene-graph-based prediction (SGPred), where we build a scene-graph-based offset regression (SGOR) mechanism to enable mutual enhancement between scene graph extraction and object localization. Thirdly, we design a cross-modal learning mechanism in SGPred. It takes scene graphs as bridges to improve the consistency between cross-modal embeddings for OV object classification. Experiments on COCO and LVIS demonstrate the effectiveness of our approach. Moreover, we show the ability of our model for OV scene graph detection, while previous OV scene graph generation methods cannot tackle this task.
Color selection plays a critical role in graphic document design and requires sufficient consideration of various contexts. However, recommending appropriate colors which harmonize with the other colors and textual contexts in documents is a challenging task, even for experienced designers. In this study, we propose a multimodal masked color model that integrates both color and textual contexts to provide text-aware color recommendation for graphic documents. Our proposed model comprises self-attention networks to capture the relationships between colors in multiple palettes, and cross-attention networks that incorporate both color and CLIP-based text representations. Our proposed method primarily focuses on color palette completion, which recommends colors based on the given colors and text. Additionally, it is applicable for another color recommendation task, full palette generation, which generates a complete color palette corresponding to the given text. Experimental results demonstrate that our proposed approach surpasses previous color palette completion methods on accuracy, color distribution, and user experience, as well as full palette generation methods concerning color diversity and similarity to the ground truth palettes.