Large-scale image-text pre-trained models enable zero-shot classification and provide consistent accuracy across various data distributions. Nonetheless, optimizing these models in downstream tasks typically requires fine-tuning, which reduces generalization to out-of-distribution (OOD) data and demands extensive computational resources. We introduce Robust Adapter (R-Adapter), a novel method for fine-tuning zero-shot models to downstream tasks while simultaneously addressing both these issues. Our method integrates lightweight modules into the pre-trained model and employs novel self-ensemble techniques to boost OOD robustness and reduce storage expenses substantially. Furthermore, we propose MPM-NCE loss designed for fine-tuning on vision-language downstream tasks. It ensures precise alignment of multiple image-text pairs and discriminative feature learning. By extending the benchmark for robust fine-tuning beyond classification to include diverse tasks such as cross-modal retrieval and open vocabulary segmentation, we demonstrate the broad applicability of R-Adapter. Our extensive experiments demonstrate that R-Adapter achieves state-of-the-art performance across a diverse set of tasks, tuning only 13% of the parameters of the CLIP encoders.
Implementing cross-modal hashing between 2D images and 3D point-cloud data is a growing concern in real-world retrieval systems. Simply applying existing cross-modal approaches to this new task fails to adequately capture latent multi-modal semantics and effectively bridge the modality gap between 2D and 3D. To address these issues without relying on hand-crafted labels, we propose contrastive masked autoencoders based self-supervised hashing (CMAH) for retrieval between images and point-cloud data. We start by contrasting 2D-3D pairs and explicitly constraining them into a joint Hamming space. This contrastive learning process ensures robust discriminability for the generated hash codes and effectively reduces the modality gap. Moreover, we utilize multi-modal auto-encoders to enhance the model's understanding of multi-modal semantics. By completing the masked image/point-cloud data modeling task, the model is encouraged to capture more localized clues. In addition, the proposed multi-modal fusion block facilitates fine-grained interactions among different modalities. Extensive experiments on three public datasets demonstrate that the proposed CMAH significantly outperforms all baseline methods.
Cross-modal retrieval is crucial in understanding latent correspondences across modalities. However, existing methods implicitly assume well-matched training data, which is impractical as real-world data inevitably involves imperfect alignments, i.e., noisy correspondences. Although some works explore similarity-based strategies to address such noise, they suffer from sub-optimal similarity predictions influenced by modality-exclusive information (MEI), e.g., background noise in images and abstract definitions in texts. This issue arises as MEI is not shared across modalities, thus aligning it in training can markedly mislead similarity predictions. Moreover, although intuitive, directly applying previous cross-modal disentanglement methods suffers from limited noise tolerance and disentanglement efficacy. Inspired by the robustness of information bottlenecks against noise, we introduce DisNCL, a novel information-theoretic framework for feature Disentanglement in Noisy Correspondence Learning, to adaptively balance the extraction of MII and MEI with certifiable optimal cross-modal disentanglement efficacy. DisNCL then enhances similarity predictions in modality-invariant subspace, thereby greatly boosting similarity-based alleviation strategy for noisy correspondences. Furthermore, DisNCL introduces soft matching targets to model noisy many-to-many relationships inherent in multi-modal input for noise-robust and accurate cross-modal alignment. Extensive experiments confirm DisNCL's efficacy by 2% average recall improvement. Mutual information estimation and visualization results show that DisNCL learns meaningful MII/MEI subspaces, validating our theoretical analyses.
The burgeoning short video industry has accelerated the advancement of video-music retrieval technology, assisting content creators in selecting appropriate music for their videos. In self-supervised training for video-to-music retrieval, the video and music samples in the dataset are separated from the same video work, so they are all one-to-one matches. This does not match the real situation. In reality, a video can use different music as background music, and a music can be used as background music for different videos. Many videos and music that are not in a pair may be compatible, leading to false negative noise in the dataset. A novel inter-intra modal (II) loss is proposed as a solution. By reducing the variation of feature distribution within the two modalities before and after the encoder, II loss can reduce the model's overfitting to such noise without removing it in a costly and laborious way. The video-music retrieval framework, II-CLVM (Contrastive Learning for Video-Music Retrieval), incorporating the II Loss, achieves state-of-the-art performance on the YouTube8M dataset. The framework II-CLVTM shows better performance when retrieving music using multi-modal video information (such as text in videos). Experiments are designed to show that II loss can effectively alleviate the problem of false negative noise in retrieval tasks. Experiments also show that II loss improves various self-supervised and supervised uni-modal and cross-modal retrieval tasks, and can obtain good retrieval models with a small amount of training samples.
With the recent burst of 2D and 3D data, cross-modal retrieval has attracted increasing attention recently. However, manual labeling by non-experts will inevitably introduce corrupted annotations given ambiguous 2D/3D content. Though previous works have addressed this issue by designing a naive division strategy with hand-crafted thresholds, their performance generally exhibits great sensitivity to the threshold value. Besides, they fail to fully utilize the valuable supervisory signals within each divided subset. To tackle this problem, we propose a Divide-and-conquer 2D-3D cross-modal Alignment and Correction framework (DAC), which comprises Multimodal Dynamic Division (MDD) and Adaptive Alignment and Correction (AAC). Specifically, the former performs accurate sample division by adaptive credibility modeling for each sample based on the compensation information within multimodal loss distribution. Then in AAC, samples in distinct subsets are exploited with different alignment strategies to fully enhance the semantic compactness and meanwhile alleviate over-fitting to noisy labels, where a self-correction strategy is introduced to improve the quality of representation. Moreover. To evaluate the effectiveness in real-world scenarios, we introduce a challenging noisy benchmark, namely Objaverse-N200, which comprises 200k-level samples annotated with 1156 realistic noisy labels. Extensive experiments on both traditional and the newly proposed benchmarks demonstrate the generality and superiority of our DAC, where DAC outperforms state-of-the-art models by a large margin. (i.e., with +5.9% gain on ModelNet40 and +5.8% on Objaverse-N200).
Visual-Language Alignment (VLA) has gained a lot of attention since CLIP's groundbreaking work. Although CLIP performs well, the typical direct latent feature alignment lacks clarity in its representation and similarity scores. On the other hand, lexical representation, a vector whose element represents the similarity between the sample and a word from the vocabulary, is a natural sparse representation and interpretable, providing exact matches for individual words. However, lexical representations are difficult to learn due to no ground-truth supervision and false-discovery issues, and thus requires complex design to train effectively. In this paper, we introduce LexVLA, a more interpretable VLA framework by learning a unified lexical representation for both modalities without complex design. We use DINOv2 as our visual model for its local-inclined features and Llama 2, a generative language model, to leverage its in-context lexical prediction ability. To avoid the false discovery, we propose an overuse penalty to refrain the lexical representation from falsely frequently activating meaningless words. We demonstrate that these two pre-trained uni-modal models can be well-aligned by fine-tuning on the modest multi-modal dataset and avoid intricate training configurations. On cross-modal retrieval benchmarks, LexVLA, trained on the CC-12M multi-modal dataset, outperforms baselines fine-tuned on larger datasets (e.g., YFCC15M) and those trained from scratch on even bigger datasets (e.g., 1.1B data, including CC-12M). We conduct extensive experiments to analyze LexVLA. Codes are available at https://github.com/Clementine24/LexVLA.
Text-to-image retrieval is a fundamental task in multimedia processing, aiming to retrieve semantically relevant cross-modal content. Traditional studies have typically approached this task as a discriminative problem, matching the text and image via the cross-attention mechanism (one-tower framework) or in a common embedding space (two-tower framework). Recently, generative cross-modal retrieval has emerged as a new research line, which assigns images with unique string identifiers and generates the target identifier as the retrieval target. Despite its great potential, existing generative approaches are limited due to the following issues: insufficient visual information in identifiers, misalignment with high-level semantics, and learning gap towards the retrieval target. To address the above issues, we propose an autoregressive voken generation method, named AVG. AVG tokenizes images into vokens, i.e., visual tokens, and innovatively formulates the text-to-image retrieval task as a token-to-voken generation problem. AVG discretizes an image into a sequence of vokens as the identifier of the image, while maintaining the alignment with both the visual information and high-level semantics of the image. Additionally, to bridge the learning gap between generative training and the retrieval target, we incorporate discriminative training to modify the learning direction during token-to-voken training. Extensive experiments demonstrate that AVG achieves superior results in both effectiveness and efficiency.
Recent studies on learning-based sound source localization have mainly focused on the localization performance perspective. However, prior work and existing benchmarks overlook a crucial aspect: cross-modal interaction, which is essential for interactive sound source localization. Cross-modal interaction is vital for understanding semantically matched or mismatched audio-visual events, such as silent objects or off-screen sounds. In this paper, we first comprehensively examine the cross-modal interaction of existing methods, benchmarks, evaluation metrics, and cross-modal understanding tasks. Then, we identify the limitations of previous studies and make several contributions to overcome the limitations. First, we introduce a new synthetic benchmark for interactive sound source localization. Second, we introduce new evaluation metrics to rigorously assess sound source localization methods, focusing on accurately evaluating both localization performance and cross-modal interaction ability. Third, we propose a learning framework with a cross-modal alignment strategy to enhance cross-modal interaction. Lastly, we evaluate both interactive sound source localization and auxiliary cross-modal retrieval tasks together to thoroughly assess cross-modal interaction capabilities and benchmark competing methods. Our new benchmarks and evaluation metrics reveal previously overlooked issues in sound source localization studies. Our proposed novel method, with enhanced cross-modal alignment, shows superior sound source localization performance. This work provides the most comprehensive analysis of sound source localization to date, with extensive validation of competing methods on both existing and new benchmarks using new and standard evaluation metrics.
Multi-modal embeddings form the foundation for vision-language models, such as CLIP embeddings, the most widely used text-image embeddings. However, these embeddings are vulnerable to subtle misalignment of cross-modal features, resulting in decreased model performance and diminished generalization. To address this problem, we design ModalChorus, an interactive system for visual probing and alignment of multi-modal embeddings. ModalChorus primarily offers a two-stage process: 1) embedding probing with Modal Fusion Map (MFM), a novel parametric dimensionality reduction method that integrates both metric and nonmetric objectives to enhance modality fusion; and 2) embedding alignment that allows users to interactively articulate intentions for both point-set and set-set alignments. Quantitative and qualitative comparisons for CLIP embeddings with existing dimensionality reduction (e.g., t-SNE and MDS) and data fusion (e.g., data context map) methods demonstrate the advantages of MFM in showcasing cross-modal features over common vision-language datasets. Case studies reveal that ModalChorus can facilitate intuitive discovery of misalignment and efficient re-alignment in scenarios ranging from zero-shot classification to cross-modal retrieval and generation.
While advancements in Vision Language Models (VLMs) have significantly improved the alignment of visual and textual data, these models primarily focus on aligning images with short descriptive captions. This focus limits their ability to handle complex text interactions, particularly with longer texts such as lengthy captions or documents, which have not been extensively explored yet. In this paper, we introduce Meet At The Embedding (MATE), a novel approach that combines the capabilities of VLMs with Large Language Models (LLMs) to overcome this challenge without the need for additional image-long text pairs. Specifically, we replace the text encoder of the VLM with a pretrained LLM-based encoder that excels in understanding long texts. To bridge the gap between VLM and LLM, MATE incorporates a projection module that is trained in a multi-stage manner. It starts by aligning the embeddings from the VLM text encoder with those from the LLM using extensive text pairs. This module is then employed to seamlessly align image embeddings closely with LLM embeddings. We propose two new cross-modal retrieval benchmarks to assess the task of connecting images with long texts (lengthy captions / documents). Extensive experimental results demonstrate that MATE effectively connects images with long texts, uncovering diverse semantic relationships.
In this paper, we present our solution for the WSDM2023 Toloka Visual Question Answering Challenge. Inspired by the application of multimodal pre-trained models to various downstream tasks(e.g., visual question answering, visual grounding, and cross-modal retrieval), we approached this competition as a visual grounding task, where the input is an image and a question, guiding the model to answer the question and display the answer as a bounding box on the image. We designed a three-stage solution for this task. Specifically, we used the visual-language pre-trained model OFA as the foundation. In the first stage, we constructed a large-scale synthetic dataset similar to the competition dataset and coarse-tuned the model to learn generalized semantic information. In the second stage, we treated the competition task as a visual grounding task, loaded the weights from the previous stage, and continued to fine-tune the model on the competition dataset, transferring the semantic information learned in the first stage to the competition task. Finally, we designed a bounding box matching and replacing post-processing strategy to correct the model's prediction results. Our team achieved a score of 76.342 on the final leaderboard, ranking second.
In the field of vision-language contrastive learning, models such as CLIP capitalize on matched image-caption pairs as positive examples and leverage within-batch non-matching pairs as negatives. This approach has led to remarkable outcomes in zero-shot image classification, cross-modal retrieval, and linear evaluation tasks. We show that the zero-shot classification and retrieval capabilities of CLIP-like models can be improved significantly through the introduction of semantically composite examples during pretraining. Inspired by CutMix in vision categorization, we create semantically composite image-caption pairs by merging elements from two distinct instances in the dataset via a novel procedure. Our method fuses the captions and blends 50% of each image to form a new composite sample. This simple technique (termed CLIP-C for CLIP Compositions), devoid of any additional computational overhead or increase in model parameters, significantly improves zero-shot image classification and cross-modal retrieval. The benefits of CLIP-C are particularly pronounced in settings with relatively limited pretraining data.
Cross-lingual Cross-modal Retrieval (CCR) is an essential task in web search, which aims to break the barriers between modality and language simultaneously and achieves image-text retrieval in the multi-lingual scenario with a single model. In recent years, excellent progress has been made based on cross-lingual cross-modal pre-training; particularly, the methods based on contrastive learning on large-scale data have significantly improved retrieval tasks. However, these methods directly follow the existing pre-training methods in the cross-lingual or cross-modal domain, leading to two problems of inconsistency in CCR: The methods with cross-lingual style suffer from the intra-modal error propagation, resulting in inconsistent recall performance across languages in the whole dataset. The methods with cross-modal style suffer from the inter-modal optimization direction bias, resulting in inconsistent rank across languages within each instance, which cannot be reflected by Recall@K. To solve these problems, we propose a simple but effective 1-to-K contrastive learning method, which treats each language equally and eliminates error propagation and optimization bias. In addition, we propose a new evaluation metric, Mean Rank Variance (MRV), to reflect the rank inconsistency across languages within each instance. Extensive experiments on four CCR datasets show that our method improves both recall rates and MRV with smaller-scale pre-trained data, achieving the new state-of-art.
Generative retrieval, which has demonstrated effectiveness in text-to-text retrieval, utilizes a sequence-to-sequence model to directly generate candidate identifiers based on natural language queries. Without explicitly computing the similarity between queries and candidates, generative retrieval surpasses dual-tower models in both speed and accuracy on large-scale corpora, providing new insights for cross-modal retrieval. However, constructing identifiers for multimodal data remains an untapped problem, and the modality gap between natural language queries and multimodal candidates hinders retrieval performance due to the absence of additional encoders. To this end, we propose a pioneering generAtive Cross-modal rEtrieval framework (ACE), which is a comprehensive framework for end-to-end cross-modal retrieval based on coarse-to-fine semantic modeling. We propose combining K-Means and RQ-VAE to construct coarse and fine tokens, serving as identifiers for multimodal data. Correspondingly, we design the coarse-to-fine feature fusion strategy to efficiently align natural language queries and candidate identifiers. ACE is the first work to comprehensively demonstrate the feasibility of generative approach on text-to-image/audio/video retrieval, challenging the dominance of the embedding-based dual-tower architecture. Extensive experiments show that ACE achieves state-of-the-art performance in cross-modal retrieval and outperforms the strong baselines on Recall@1 by 15.27% on average.
By leveraging the kernel trick in the output space, kernel-induced losses provide a principled way to define structured output prediction tasks for a wide variety of output modalities. In particular, they have been successfully used in the context of surrogate non-parametric regression, where the kernel trick is typically exploited in the input space as well. However, when inputs are images or texts, more expressive models such as deep neural networks seem more suited than non-parametric methods. In this work, we tackle the question of how to train neural networks to solve structured output prediction tasks, while still benefiting from the versatility and relevance of kernel-induced losses. We design a novel family of deep neural architectures, whose last layer predicts in a data-dependent finite-dimensional subspace of the infinite-dimensional output feature space deriving from the kernel-induced loss. This subspace is chosen as the span of the eigenfunctions of a randomly-approximated version of the empirical kernel covariance operator. Interestingly, this approach unlocks the use of gradient descent algorithms (and consequently of any neural architecture) for structured prediction. Experiments on synthetic tasks as well as real-world supervised graph prediction problems show the relevance of our method.
Web-crawled image-text pairs are inherently noisy. Prior studies demonstrate that semantically aligning and enriching textual descriptions of these pairs can significantly enhance model training across various vision-language tasks, particularly text-to-image generation. However, large-scale investigations in this area remain predominantly closed-source. Our paper aims to bridge this community effort, leveraging the powerful and \textit{open-sourced} LLaMA-3, a GPT-4 level LLM. Our recaptioning pipeline is simple: first, we fine-tune a LLaMA-3-8B powered LLaVA-1.5 and then employ it to recaption 1.3 billion images from the DataComp-1B dataset. Our empirical results confirm that this enhanced dataset, Recap-DataComp-1B, offers substantial benefits in training advanced vision-language models. For discriminative models like CLIP, we observe enhanced zero-shot performance in cross-modal retrieval tasks. For generative models like text-to-image Diffusion Transformers, the generated images exhibit a significant improvement in alignment with users' text instructions, especially in following complex queries. Our project page is https://www.haqtu.me/Recap-Datacomp-1B/
We present DenseAV, a novel dual encoder grounding architecture that learns high-resolution, semantically meaningful, and audio-visually aligned features solely through watching videos. We show that DenseAV can discover the ``meaning'' of words and the ``location'' of sounds without explicit localization supervision. Furthermore, it automatically discovers and distinguishes between these two types of associations without supervision. We show that DenseAV's localization abilities arise from a new multi-head feature aggregation operator that directly compares dense image and audio representations for contrastive learning. In contrast, many other systems that learn ``global'' audio and video representations cannot localize words and sound. Finally, we contribute two new datasets to improve the evaluation of AV representations through speech and sound prompted semantic segmentation. On these and other datasets we show DenseAV dramatically outperforms the prior art on speech and sound prompted semantic segmentation. DenseAV outperforms the previous state-of-the-art, ImageBind, on cross-modal retrieval using fewer than half of the parameters. Project Page: \href{https://aka.ms/denseav}{https://aka.ms/denseav}
Over 85 million computed tomography (CT) scans are performed annually in the US, of which approximately one quarter focus on the abdomen. Given the current radiologist shortage, there is a large impetus to use artificial intelligence to alleviate the burden of interpreting these complex imaging studies. Prior state-of-the-art approaches for automated medical image interpretation leverage vision language models (VLMs). However, current medical VLMs are generally limited to 2D images and short reports, and do not leverage electronic health record (EHR) data for supervision. We introduce Merlin - a 3D VLM that we train using paired CT scans (6+ million images from 15,331 CTs), EHR diagnosis codes (1.8+ million codes), and radiology reports (6+ million tokens). We evaluate Merlin on 6 task types and 752 individual tasks. The non-adapted (off-the-shelf) tasks include zero-shot findings classification (31 findings), phenotype classification (692 phenotypes), and zero-shot cross-modal retrieval (image to findings and image to impressions), while model adapted tasks include 5-year disease prediction (6 diseases), radiology report generation, and 3D semantic segmentation (20 organs). We perform internal validation on a test set of 5,137 CTs, and external validation on 7,000 clinical CTs and on two public CT datasets (VerSe, TotalSegmentator). Beyond these clinically-relevant evaluations, we assess the efficacy of various network architectures and training strategies to depict that Merlin has favorable performance to existing task-specific baselines. We derive data scaling laws to empirically assess training data needs for requisite downstream task performance. Furthermore, unlike conventional VLMs that require hundreds of GPUs for training, we perform all training on a single GPU.
In this paper, we study a challenging but less-touched problem in cross-modal retrieval, i.e., partially mismatched pairs (PMPs). Specifically, in real-world scenarios, a huge number of multimedia data (e.g., the Conceptual Captions dataset) are collected from the Internet, and thus it is inevitable to wrongly treat some irrelevant cross-modal pairs as matched. Undoubtedly, such a PMP problem will remarkably degrade the cross-modal retrieval performance. To tackle this problem, we derive a unified theoretical Robust Cross-modal Learning framework (RCL) with an unbiased estimator of the cross-modal retrieval risk, which aims to endow the cross-modal retrieval methods with robustness against PMPs. In detail, our RCL adopts a novel complementary contrastive learning paradigm to address the following two challenges, i.e., the overfitting and underfitting issues. On the one hand, our method only utilizes the negative information which is much less likely false compared with the positive information, thus avoiding the overfitting issue to PMPs. However, these robust strategies could induce underfitting issues, thus making training models more difficult. On the other hand, to address the underfitting issue brought by weak supervision, we present to leverage of all available negative pairs to enhance the supervision contained in the negative information. Moreover, to further improve the performance, we propose to minimize the upper bounds of the risk to pay more attention to hard samples. To verify the effectiveness and robustness of the proposed method, we carry out comprehensive experiments on five widely-used benchmark datasets compared with nine state-of-the-art approaches w.r.t. the image-text and video-text retrieval tasks. The code is available at https://github.com/penghu-cs/RCL .
The success of existing cross-modal retrieval (CMR) methods heavily rely on the assumption that the annotated cross-modal correspondence is faultless. In practice, however, the correspondence of some pairs would be inevitably contaminated during data collection or annotation, thus leading to the so-called Noisy Correspondence (NC) problem. To alleviate the influence of NC, we propose a novel method termed Consistency REfining And Mining (CREAM) by revealing and exploiting the difference between correspondence and consistency. Specifically, the correspondence and the consistency only be coincident for true positive and true negative pairs, while being distinct for false positive and false negative pairs. Based on the observation, CREAM employs a collaborative learning paradigm to detect and rectify the correspondence of positives, and a negative mining approach to explore and utilize the consistency. Thanks to the consistency refining and mining strategy of CREAM, the overfitting on the false positives could be prevented and the consistency rooted in the false negatives could be exploited, thus leading to a robust CMR method. Extensive experiments verify the effectiveness of our method on three image-text benchmarks including Flickr30K, MS-COCO, and Conceptual Captions. Furthermore, we adopt our method into the graph matching task and the results demonstrate the robustness of our method against fine-grained NC problem. The code is available on https://github.com/XLearning-SCU/2024-TIP-CREAM .