Contrastive Language-Image Pretraining (CLIP) has demonstrated impressive generalization on vision-language tasks by aligning images and short texts. However, its inherent 77-token length limits the capacity of capturing complex semantics in long captions. Existing long-text adaptations for CLIP typically rely on either multi-stage training or truncation-based alignment, both inevitably resulting in semantic degradation and cumbersome tuning. Therefore, we propose OneLIP, a unified framework that extends CLIP to understand long captions within a single training stage, eliminating the need for brittle truncation or multi-stage pipelines. OneLIP addresses semantic degradation by introducing two key innovations: (1) Token Refinement and Importance-guided Modeling (TRIM) module, which selects and refines informative tokens via SVD-based contribution scoring and cross-modal relevance modeling; (2) Per-sample Online Hard Negative Mining (PO-HNM) strategy dynamically maintains sample-specific negatives based on dual-consistency difficulty tracking, which is superior in long-text scenarios where key semantics are distributed in scattered positions. Extensive experiments on long-text image retrieval, short-text image retrieval, zero-shot classification, and text-to-image generation demonstrate OneLIP's robustness and versatility across diverse input lengths, offering a faithful solution for long-text representation learning of CLIP.

Previous research on sparse feature matching typically involves a staged optimization process of keypoint detection, description, and matching. While it allows the network to adapt to specific inputs, it may limit the network's expressive capability and the overall architectural flexibility. In this study, we rethink the matching framework and propose to directly match any given keypoints, optimizing the matching network in an approximately end-to-end manner. To achieve this, firstly, we dynamically sample random positions within the images as assumed keypoints during training, allowing the network to explore a broader matching space. Secondly, we replace specific descriptors with high-efficiency sparse embeddings at multi levels of the image, facilitating the direct learning of underlying textures. Thirdly, we propose a novel and promising architecture, called Proposal-Guided TRansformer (PGTR), which aggregates context information from neighboring match proposals instead of searching globally with local features. PGTR works especially well under our training approach, and attain a synergistic advantage in terms of performance and efficiency. The overall pipeline achieves outstanding performance on various keypoints without any retraining, and can be flexibly reused when new keypoints emerge, making it valuable for real-world applications. Code will be available.

’Pre-training + fine-tuning’ has been widely used in various downstream tasks. Parameter-efficient fine-tuning (PEFT) has demonstrated higher efficiency and promising performance compared to traditional full-tuning. The widely used adapter-based and prompt-based methods in PEFT can be uniformly represented as adding an MLP structure to the pre-trained model. These methods are prone to over-fitting in downstream tasks, due to the difference in data scale and distribution. To address this issue, we propose a new adapter-based PEFT module, i.e., LoKi, which consists of an encoder, a learnable activation layer, and a decoder. To maintain the simplicity of LoKi, we use single-layer linear networks for the encoder and decoder, and for the learnable activation layer, we use a Kolmogorov-Arnold Network (KAN) with the minimal number of layers (only 2 KAN linear layers). With a bottleneck rate much lower than that of Adapter, LoKi is equipped with fewer parameters (only half of Adapter) and eliminates slow training speed and high memory usage of KAN. We conduct extensive experiments on LoKi under image classification and video action recognition across 9 datasets. LoKi demonstrates highly competitive generalization performance compared to other PEFT methods with fewer tunable parameters, ensuring both effectiveness and efficiency.

In this paper, we identify a common yet challenging issue in image-text matching, i.e., clone negatives: negative image-text pairs that semantically resemble positive pairs, leading to ambiguous and sub-optimal matching outcomes. To tackle this issue, we propose Adaptive Contrastive Learning (AdaCL), which introduces two margin parameters along with a modulating anchor to dynamically strengthen the compactness between positives and mitigate the influence of clone negatives. The modulating anchor is selected based on the distribution of negative samples without the need for explicit training, allowing for progressive tuning and advanced in-batch supervision. Extensive experiments across several tasks demonstrate the effectiveness of AdaCL in image-text matching. Furthermore, we extend AdaCL to weakly-supervised image-text matching by replacing human-annotated descriptions with automatically generated captions, thereby increasing the number of potential clone negatives. AdaCL maintains robustness in this setting, alleviating the reliance on crowd-sourced annotations and laying a foundation for scalable vision-language contrastive learning.

Current methods for image-text retrieval commonly propose various fusion modules to achieve robust visual-textual alignment, primarily relying on in-batch learning to guide the matching process. Some follow-up methods seek to enlarge the number of negative samples to boost image-text contrastive learning. However, these methods often face challenges posed by semantic-consistent negatives, i.e., negative samples that share correspondence with the ground truth, leading to confusion in learning cross-modal semantics. To address this issue, we propose a novel Retrieve with Authentic Negative Repository Learning (ReAL) method, which constructs a specific Authentic Negative Repository filled with valuable negative sample pairs. By introducing a Unique Negative Filter with a Discriminative Triplet Ranking Loss, ReAL effectively filters out the semantic-consistent negatives through similarity distribution analysis and threshold learning. Moreover, existing fusion paradigms suffer from intricate use of fine-grained representations from word- and region-level instances to progressively refine the fused embedding. In this article, we propose a lightweight Cluster Refinement Module to exploit cross-modal semantics in a 1-way-1-out paradigm. Each visual-textual alignment can spontaneously uncover correlations with adjacent alignments through aggregation and re-allocation, without the need for a redundant and cost-inefficient refinement stage. Furthermore, ReAL employs dual momentum encoders with two memory banks, expanding the selection range of the Authentic Negative Repository to include a broader set of negatives. Extensive experiments conducted on Flickr30K, MS-COCO, and the augmented Flickr30K (with more hard negatives) demonstrate the superiority and robustness of ReAL, while also showcasing its significantly reduced inference time compared to other competitive baselines.

Finding correspondences between images is essential for many computer vision tasks and sparse matching pipelines have been popular for decades. However, matching noise within and between images, along with inconsistent key-point detection, frequently degrades the matching performance. We review these problems and thus propose: 1) a novel and unified Filtering and Calibrating (FC) approach that jointly rejects outliers and optimizes inliers, and 2) leveraging both the matching context and the underlying image texture to remove matching uncertainties. Under the guidance of the above innovations, we construct Filtering and Calibrating Graph Neural Network (FC-GNN), which follows the FC approach to recover reliable and accurate correspondences from various interferences. FC-GNN conducts an effectively combined inference of contextual and local information through careful embedding and multiple information aggregations, predicting confidence scores and calibration offsets for the input correspondences to jointly filter out outliers and improve pixel-level matching accuracy. Moreover, we exploit the local coherence of matches to perform inference on local graphs, thereby reducing computational complexity. Overall, FC-GNN operates at lightning speed and can greatly boost the performance of diverse matching pipelines across various tasks, showcasing the immense potential of such approaches to become standard and pivotal components of image matching. Code is avaiable at https://github.com/xuy123456/fcgnn.

Human-Object Interaction (HOI) detection aims to localize human-object pairs and comprehend their interactions. Recently, two-stage transformer-based methods have demonstrated competitive performance. However, these methods frequently focus on object appearance features and ignore global contextual information. Besides, vision-language model CLIP which effectively aligns visual and text embeddings has shown great potential in zero-shot HOI detection. Based on the former facts, We introduce a novel HOI detector named ISA-HOI, which extensively leverages knowledge from CLIP, aligning interactive semantics between visual and textual features. We first extract global context of image and local features of object to Improve interaction Features in images (IF). On the other hand, we propose a Verb Semantic Improvement (VSI) module to enhance textual features of verb labels via cross-modal fusion. Ultimately, our method achieves competitive results on the HICO-DET and V-COCO benchmarks with much fewer training epochs, and outperforms the state-of-the-art under zero-shot settings.

Multimodal understanding aims at constructing semantic correlations among modalities of data while performing various downstream tasks. As one of the primary multimodal downstream tasks, image-text retrieval imposes a high demand on semantic alignment because of the independent expression paradigms of images and text. Existing methods mainly construct a joint embedding space at a single granularity level (either global or local). However, such single reasoning paradigms lack granularity interaction, resulting in semantic inconsistency and cross-domain catastrophes. To address these issues, we design a novel Joint Intra and Inter-grained Network (JIIGNet), focusing on not only intra- but also inter-grained interaction between modalities by combining scene information (global) with region-level (local) instances. Specifically, we simultaneously initiate three specific alignment modules, i.e., global-grained, local-grained, and cross-grained alignment modules, followed by Triplet Attention Refinement to better refine the fused embedding at the alignment-level with proper self and cross attention. For different scenarios, a Style Adaptation Head is further designed to smartly accommodate different samples. We validate JIIGNet through extensive experiments conducted on two widely used datasets: Flickr-30 K and MS-COCO, demonstrating the effectiveness of our proposed method.

Social interaction is a common phenomenon in human societies. Different from discovering groups based on the similarity of individuals’ actions, social interaction focuses more on the mutual influence between people. Although people can easily judge whether or not there are social interactions in a real-world scene, it is difficult for an intelligent system to discover social interactions. Initiating and concluding social interactions are greatly influenced by an individual’s social cognition and the surrounding environment, which are closely related to psychology. Thus, converting the psychological factors that impact social interactions into quantifiable visual representations and creating a model for interaction relationships poses a significant challenge. To this end, we propose a Psychology-Guided Environment Aware Network (PEAN) that models social interaction among people in videos using supervised learning. Specifically, we divide the surrounding environment into scene-aware visual-based and human-aware visual-based descriptions. For the scene-aware visual clue, we utilize 3D features as global visual representations. For the human-aware visual clue, we consider instance-based location and behaviour-related visual representations to map human-centred interaction elements in social psychology: distance, openness, and orientation. In addition, we design an environment aware mechanism to integrate features from visual clues, with a Transformer to explore the relation between individuals and construct pairwise interaction strength features. The interaction intensity matrix reflecting the mutual nature of the interaction is obtained by processing the interaction strength features with the interaction discovery module. An interaction constrained loss function composed of interaction critical loss function and smoothFβloss function is proposed to optimize the whole framework to improve the distinction of the interaction matrix and alleviate class imbalance caused by pairwise interaction sparsity. Given the diversity of real-world interactions, we collect a new dataset named Social Basketball Activity Dataset (Soical-BAD), covering complex social interactions. Our method achieves the best performance among social-CAD, social-BAD, and their combined dataset named Video Social Interaction Dataset (VSID).

Text-to-image person search is challenging due to the cross-scale correspondences and information inequality between modalities. Specifically, images and text are complexly linked at different scales and images are usually more informative and complete than text. It is crucial to establish semantic correlations between modalities and focus on task-relevant information in images. In this paper, we propose a novel Adaptive and Collaborative Multi-scale Alignment network (ACMA) for text-based person search that learns semantically consistent and information-aligned multi-modal representations. Firstly, we introduce a novel joint embedding module that adaptively integrates features of different pixels and words, thereby extracting semantically consistent multi-modal features at different scales. Second, we design a cross-modal fusion feature-based auxiliary visual branch to guide the extraction of key visual features that are beneficial for cross-modal matching. Extensive experiments validate that ACMA outperforms the state-of-the-art method.