Xiatian Zhu

Junting Pan, Ziyi Lin, Xiatian Zhu et al.

Capitalizing on large pre-trained models for various downstream tasks of interest have recently emerged with promising performance. Due to the ever-growing model size, the standard full fine-tuning based task adaptation strategy becomes prohibitively costly in terms of model training and storage. This has led to a new research direction in parameter-efficient transfer learning. However, existing attempts typically focus on downstream tasks from the same modality (e.g., image understanding) of the pre-trained model. This creates a limit because in some specific modalities, (e.g., video understanding) such a strong pre-trained model with sufficient knowledge is less or not available. In this work, we investigate such a novel cross-modality transfer learning setting, namely parameter-efficient image-to-video transfer learning. To solve this problem, we propose a new Spatio-Temporal Adapter (ST-Adapter) for parameter-efficient fine-tuning per video task. With a built-in spatio-temporal reasoning capability in a compact design, ST-Adapter enables a pre-trained image model without temporal knowledge to reason about dynamic video content at a small (~8%) per-task parameter cost, requiring approximately 20 times fewer updated parameters compared to previous work. Extensive experiments on video action recognition tasks show that our ST-Adapter can match or even outperform the strong full fine-tuning strategy and state-of-the-art video models, whilst enjoying the advantage of parameter efficiency. The code and model are available at https://github.com/linziyi96/st-adapter

Junting Pan, Adrian Bulat, Fuwen Tan et al.

Self-attention based models such as vision transformers (ViTs) have emerged as a very competitive architecture alternative to convolutional neural networks (CNNs) in computer vision. Despite increasingly stronger variants with ever-higher recognition accuracies, due to the quadratic complexity of self-attention, existing ViTs are typically demanding in computation and model size. Although several successful design choices (e.g., the convolutions and hierarchical multi-stage structure) of prior CNNs have been reintroduced into recent ViTs, they are still not sufficient to meet the limited resource requirements of mobile devices. This motivates a very recent attempt to develop light ViTs based on the state-of-the-art MobileNet-v2, but still leaves a performance gap behind. In this work, pushing further along this under-studied direction we introduce EdgeViTs, a new family of light-weight ViTs that, for the first time, enable attention-based vision models to compete with the best light-weight CNNs in the tradeoff between accuracy and on-device efficiency. This is realized by introducing a highly cost-effective local-global-local (LGL) information exchange bottleneck based on optimal integration of self-attention and convolutions. For device-dedicated evaluation, rather than relying on inaccurate proxies like the number of FLOPs or parameters, we adopt a practical approach of focusing directly on on-device latency and, for the first time, energy efficiency. Specifically, we show that our models are Pareto-optimal when both accuracy-latency and accuracy-energy trade-offs are considered, achieving strict dominance over other ViTs in almost all cases and competing with the most efficient CNNs. Code is available at https://github.com/saic-fi/edgevit.

Jing Yang, Xiatian Zhu, Adrian Bulat et al. · cambridge

Existing knowledge distillation methods mostly focus on distillation of teacher's prediction and intermediate activation. However, the structured representation, which arguably is one of the most critical ingredients of deep models, is largely overlooked. In this work, we propose a novel {\em \modelname{}} ({\bf\em \shortname{})} method dedicated for distilling representational knowledge semantically from a pretrained teacher to a target student. The key idea is that we leverage the teacher's classifier as a semantic critic for evaluating the representations of both teacher and student and distilling the semantic knowledge with high-order structured information over all feature dimensions. This is accomplished by introducing a notion of cross-network logit computed through passing student's representation into teacher's classifier. Further, considering the set of seen classes as a basis for the semantic space in a combinatorial perspective, we scale \shortname{} to unseen classes for enabling effective exploitation of largely available, arbitrary unlabeled training data. At the problem level, this establishes an interesting connection between knowledge distillation with open-set semi-supervised learning (SSL). Extensive experiments show that our \shortname{} outperforms significantly previous state-of-the-art knowledge distillation methods on both coarse object classification and fine face recognition tasks, as well as less studied yet practically crucial binary network distillation. Under more realistic open-set SSL settings we introduce, we reveal that knowledge distillation is generally more effective than existing Out-Of-Distribution (OOD) sample detection, and our proposed \shortname{} is superior over both previous distillation and SSL competitors. The source code is available at \url{https://github.com/jingyang2017/SRD\_ossl}.

Zeyu Yang, Nan Song, Wei Li et al.

Existing top-performance autonomous driving systems typically rely on the multi-modal fusion strategy for reliable scene understanding. This design is however fundamentally restricted due to overlooking the modality-specific strengths and finally hampering the model performance. To address this limitation, in this work, we introduce a novel modality interaction strategy that allows individual per-modality representations to be learned and maintained throughout, enabling their unique characteristics to be exploited during the whole perception pipeline. To demonstrate the effectiveness of the proposed strategy, we design DeepInteraction++, a multi-modal interaction framework characterized by a multi-modal representational interaction encoder and a multi-modal predictive interaction decoder. Specifically, the encoder is implemented as a dual-stream Transformer with specialized attention operation for information exchange and integration between separate modality-specific representations. Our multi-modal representational learning incorporates both object-centric, precise sampling-based feature alignment and global dense information spreading, essential for the more challenging planning task. The decoder is designed to iteratively refine the predictions by alternately aggregating information from separate representations in a unified modality-agnostic manner, realizing multi-modal predictive interaction. Extensive experiments demonstrate the superior performance of the proposed framework on both 3D object detection and end-to-end autonomous driving tasks. Our code is available at https://github.com/fudan-zvg/DeepInteraction.

Sauradip Nag, Xiatian Zhu, Yi-Zhe Song et al.

Existing temporal action detection (TAD) methods rely on large training data including segment-level annotations, limited to recognizing previously seen classes alone during inference. Collecting and annotating a large training set for each class of interest is costly and hence unscalable. Zero-shot TAD (ZS-TAD) resolves this obstacle by enabling a pre-trained model to recognize any unseen action classes. Meanwhile, ZS-TAD is also much more challenging with significantly less investigation. Inspired by the success of zero-shot image classification aided by vision-language (ViL) models such as CLIP, we aim to tackle the more complex TAD task. An intuitive method is to integrate an off-the-shelf proposal detector with CLIP style classification. However, due to the sequential localization (e.g, proposal generation) and classification design, it is prone to localization error propagation. To overcome this problem, in this paper we propose a novel zero-Shot Temporal Action detection model via Vision-LanguagE prompting (STALE). Such a novel design effectively eliminates the dependence between localization and classification by breaking the route for error propagation in-between. We further introduce an interaction mechanism between classification and localization for improved optimization. Extensive experiments on standard ZS-TAD video benchmarks show that our STALE significantly outperforms state-of-the-art alternatives. Besides, our model also yields superior results on supervised TAD over recent strong competitors. The PyTorch implementation of STALE is available at https://github.com/sauradip/STALE.

Haosen Yang, Deng Huang, Bin Wen et al.

Masked autoencoders (MAEs) have emerged recently as art self-supervised spatiotemporal representation learners. Inheriting from the image counterparts, however, existing video MAEs still focus largely on static appearance learning whilst are limited in learning dynamic temporal information hence less effective for video downstream tasks. To resolve this drawback, in this work we present a motion-aware variant -- MotionMAE. Apart from learning to reconstruct individual masked patches of video frames, our model is designed to additionally predict the corresponding motion structure information over time. This motion information is available at the temporal difference of nearby frames. As a result, our model can extract effectively both static appearance and dynamic motion spontaneously, leading to superior spatiotemporal representation learning capability. Extensive experiments show that our MotionMAE outperforms significantly both supervised learning baseline and state-of-the-art MAE alternatives, under both domain-specific and domain-generic pretraining-then-finetuning settings. In particular, when using ViT-B as the backbone our MotionMAE surpasses the prior art model by a margin of 1.2% on Something-Something V2 and 3.2% on UCF101 in domain-specific pretraining setting. Encouragingly, it also surpasses the competing MAEs by a large margin of over 3% on the challenging video object segmentation task. The code is available at https://github.com/happy-hsy/MotionMAE.

Yanbei Chen, Massimiliano Mancini, Xiatian Zhu et al. · cambridge

State-of-the-art deep learning models are often trained with a large amount of costly labeled training data. However, requiring exhaustive manual annotations may degrade the model's generalizability in the limited-label regime. Semi-supervised learning and unsupervised learning offer promising paradigms to learn from an abundance of unlabeled visual data. Recent progress in these paradigms has indicated the strong benefits of leveraging unlabeled data to improve model generalization and provide better model initialization. In this survey, we review the recent advanced deep learning algorithms on semi-supervised learning (SSL) and unsupervised learning (UL) for visual recognition from a unified perspective. To offer a holistic understanding of the state-of-the-art in these areas, we propose a unified taxonomy. We categorize existing representative SSL and UL with comprehensive and insightful analysis to highlight their design rationales in different learning scenarios and applications in different computer vision tasks. Lastly, we discuss the emerging trends and open challenges in SSL and UL to shed light on future critical research directions.

Xiao Han, Xiatian Zhu, Licheng Yu et al.

In the fashion domain, there exists a variety of vision-and-language (V+L) tasks, including cross-modal retrieval, text-guided image retrieval, multi-modal classification, and image captioning. They differ drastically in each individual input/output format and dataset size. It has been common to design a task-specific model and fine-tune it independently from a pre-trained V+L model (e.g., CLIP). This results in parameter inefficiency and inability to exploit inter-task relatedness. To address such issues, we propose a novel FAshion-focused Multi-task Efficient learning method for Vision-and-Language tasks (FAME-ViL) in this work. Compared with existing approaches, FAME-ViL applies a single model for multiple heterogeneous fashion tasks, therefore being much more parameter-efficient. It is enabled by two novel components: (1) a task-versatile architecture with cross-attention adapters and task-specific adapters integrated into a unified V+L model, and (2) a stable and effective multi-task training strategy that supports learning from heterogeneous data and prevents negative transfer. Extensive experiments on four fashion tasks show that our FAME-ViL can save 61.5% of parameters over alternatives, while significantly outperforming the conventional independently trained single-task models. Code is available at https://github.com/BrandonHanx/FAME-ViL.

Xiao Han, Licheng Yu, Xiatian Zhu et al.

Large-scale Vision-and-Language (V+L) pre-training for representation learning has proven to be effective in boosting various downstream V+L tasks. However, when it comes to the fashion domain, existing V+L methods are inadequate as they overlook the unique characteristics of both the fashion V+L data and downstream tasks. In this work, we propose a novel fashion-focused V+L representation learning framework, dubbed as FashionViL. It contains two novel fashion-specific pre-training tasks designed particularly to exploit two intrinsic attributes with fashion V+L data. First, in contrast to other domains where a V+L data point contains only a single image-text pair, there could be multiple images in the fashion domain. We thus propose a Multi-View Contrastive Learning task for pulling closer the visual representation of one image to the compositional multimodal representation of another image+text. Second, fashion text (e.g., product description) often contains rich fine-grained concepts (attributes/noun phrases). To exploit this, a Pseudo-Attributes Classification task is introduced to encourage the learned unimodal (visual/textual) representations of the same concept to be adjacent. Further, fashion V+L tasks uniquely include ones that do not conform to the common one-stream or two-stream architectures (e.g., text-guided image retrieval). We thus propose a flexible, versatile V+L model architecture consisting of a modality-agnostic Transformer so that it can be flexibly adapted to any downstream tasks. Extensive experiments show that our FashionViL achieves a new state of the art across five downstream tasks. Code is available at https://github.com/BrandonHanx/mmf.

Sauradip Nag, Xiatian Zhu, Yi-Zhe Song et al.

Existing temporal action detection (TAD) methods rely on generating an overwhelmingly large number of proposals per video. This leads to complex model designs due to proposal generation and/or per-proposal action instance evaluation and the resultant high computational cost. In this work, for the first time, we propose a proposal-free Temporal Action detection model with Global Segmentation mask (TAGS). Our core idea is to learn a global segmentation mask of each action instance jointly at the full video length. The TAGS model differs significantly from the conventional proposal-based methods by focusing on global temporal representation learning to directly detect local start and end points of action instances without proposals. Further, by modeling TAD holistically rather than locally at the individual proposal level, TAGS needs a much simpler model architecture with lower computational cost. Extensive experiments show that despite its simpler design, TAGS outperforms existing TAD methods, achieving new state-of-the-art performance on two benchmarks. Importantly, it is ~ 20x faster to train and ~1.6x more efficient for inference. Our PyTorch implementation of TAGS is available at https://github.com/sauradip/TAGS .

Sauradip Nag, Mengmeng Xu, Xiatian Zhu et al.

Few-shot (FS) and zero-shot (ZS) learning are two different approaches for scaling temporal action detection (TAD) to new classes. The former adapts a pretrained vision model to a new task represented by as few as a single video per class, whilst the latter requires no training examples by exploiting a semantic description of the new class. In this work, we introduce a new multi-modality few-shot (MMFS) TAD problem, which can be considered as a marriage of FS-TAD and ZS-TAD by leveraging few-shot support videos and new class names jointly. To tackle this problem, we further introduce a novel MUlti-modality PromPt mETa-learning (MUPPET) method. This is enabled by efficiently bridging pretrained vision and language models whilst maximally reusing already learned capacity. Concretely, we construct multi-modal prompts by mapping support videos into the textual token space of a vision-language model using a meta-learned adapter-equipped visual semantics tokenizer. To tackle large intra-class variation, we further design a query feature regulation scheme. Extensive experiments on ActivityNetv1.3 and THUMOS14 demonstrate that our MUPPET outperforms state-of-the-art alternative methods, often by a large margin. We also show that our MUPPET can be easily extended to tackle the few-shot object detection problem and again achieves the state-of-the-art performance on MS-COCO dataset. The code will be available in https://github.com/sauradip/MUPPET

Qianshan Zhan, Qian Wang, Da Li et al.

Multi-modal Domain Generalization (MMDG) seeks to leverage complementary modalities to enhance model robustness on unseen domains. Despite extensive progress in Multi-modal Learning (MML) and Domain Generalization (DG) as individual fields, their systematic integration remains under-explored. Current MMDG research is largely confined to action recognition and lacks standardized evaluation protocols. To address this, we introduce MMDG-Bench, a comprehensive benchmark featuring two foundational frameworks: DG then MML (D2M) and MML then DG (M2D). We provide unified experimental protocols across diverse tasks, including video-audio-flow action recognition and RGB-Depth-IR face anti-spoofing. By instantiating ten MMDG baselines through pairing a unified MML configuration with five DG techniques under both D2M and M2D orderings, we demonstrate that these structured combinations frequently outperform existing state-of-the-art methods, underscoring the necessity of a unified benchmarking effort. Our analysis yields three key insights: (1) Integrating DG techniques provides consistent generalization gains across various backbones, whereas non-DG methods are highly sensitive to backbone shifts; (2) The optimal framework choice depends on inter-modal stability: D2M excels when modal relations are stable across domains, while M2D is more robust to cross-domain relational variance; (3) Stronger backbones yield amplified performance dividends when integrated into our structured frameworks. MMDG-Bench provides a principled foundation and actionable design guidelines for future research in multi-modal robustness. Code is released at https://github.com/qszhan/MMDG-Bench.

Sauradip Nag, Xiatian Zhu, Jiankang Deng et al.

We propose a new formulation of temporal action detection (TAD) with denoising diffusion, DiffTAD in short. Taking as input random temporal proposals, it can yield action proposals accurately given an untrimmed long video. This presents a generative modeling perspective, against previous discriminative learning manners. This capability is achieved by first diffusing the ground-truth proposals to random ones (i.e., the forward/noising process) and then learning to reverse the noising process (i.e., the backward/denoising process). Concretely, we establish the denoising process in the Transformer decoder (e.g., DETR) by introducing a temporal location query design with faster convergence in training. We further propose a cross-step selective conditioning algorithm for inference acceleration. Extensive evaluations on ActivityNet and THUMOS show that our DiffTAD achieves top performance compared to previous art alternatives. The code will be made available at https://github.com/sauradip/DiffusionTAD.

Peng Xu, Xiatian Zhu, David A. Clifton

Transformer is a promising neural network learner, and has achieved great success in various machine learning tasks. Thanks to the recent prevalence of multimodal applications and big data, Transformer-based multimodal learning has become a hot topic in AI research. This paper presents a comprehensive survey of Transformer techniques oriented at multimodal data. The main contents of this survey include: (1) a background of multimodal learning, Transformer ecosystem, and the multimodal big data era, (2) a theoretical review of Vanilla Transformer, Vision Transformer, and multimodal Transformers, from a geometrically topological perspective, (3) a review of multimodal Transformer applications, via two important paradigms, i.e., for multimodal pretraining and for specific multimodal tasks, (4) a summary of the common challenges and designs shared by the multimodal Transformer models and applications, and (5) a discussion of open problems and potential research directions for the community.

Sauradip Nag, Xiatian Zhu, Yi-Zhe Song et al.

Existing temporal action detection (TAD) methods rely on a large number of training data with segment-level annotations. Collecting and annotating such a training set is thus highly expensive and unscalable. Semi-supervised TAD (SS-TAD) alleviates this problem by leveraging unlabeled videos freely available at scale. However, SS-TAD is also a much more challenging problem than supervised TAD, and consequently much under-studied. Prior SS-TAD methods directly combine an existing proposal-based TAD method and a SSL method. Due to their sequential localization (e.g, proposal generation) and classification design, they are prone to proposal error propagation. To overcome this limitation, in this work we propose a novel Semi-supervised Temporal action detection model based on PropOsal-free Temporal mask (SPOT) with a parallel localization (mask generation) and classification architecture. Such a novel design effectively eliminates the dependence between localization and classification by cutting off the route for error propagation in-between. We further introduce an interaction mechanism between classification and localization for prediction refinement, and a new pretext task for self-supervised model pre-training. Extensive experiments on two standard benchmarks show that our SPOT outperforms state-of-the-art alternatives, often by a large margin. The PyTorch implementation of SPOT is available at https://github.com/sauradip/SPOT

Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song et al.

This paper propels creative control in generative visual AI by allowing users to "select". Departing from traditional text or sketch-based methods, we for the first time allow users to choose visual concepts by parts for their creative endeavors. The outcome is fine-grained generation that precisely captures selected visual concepts, ensuring a holistically faithful and plausible result. To achieve this, we first parse objects into parts through unsupervised feature clustering. Then, we encode parts into text tokens and introduce an entropy-based normalized attention loss that operates on them. This loss design enables our model to learn generic prior topology knowledge about object's part composition, and further generalize to novel part compositions to ensure the generation looks holistically faithful. Lastly, we employ a bottleneck encoder to project the part tokens. This not only enhances fidelity but also accelerates learning, by leveraging shared knowledge and facilitating information exchange among instances. Visual results in the paper and supplementary material showcase the compelling power of PartCraft in crafting highly customized, innovative creations, exemplified by the "charming" and creative birds. Code is released at https://github.com/kamwoh/partcraft.

Jiachen Lu, Junge Zhang, Xiatian Zhu et al.

Vision transformers (ViTs) have pushed the state-of-the-art for visual perception tasks. The self-attention mechanism underpinning the strength of ViTs has a quadratic complexity in both computation and memory usage. This motivates the development of approximating the self-attention at linear complexity. However, an in-depth analysis in this work reveals that existing methods are either theoretically flawed or empirically ineffective for visual recognition. We identify that their limitations are rooted in the inheritance of softmax-based self-attention during approximations, that is, normalizing the scaled dot-product between token feature vectors using the softmax function. As preserving the softmax operation challenges any subsequent linearization efforts. By this insight, a family of Softmax-Free Transformers (SOFT) are proposed. Specifically, a Gaussian kernel function is adopted to replace the dot-product similarity, enabling a full self-attention matrix to be approximated under low-rank matrix decomposition. For computational robustness, we estimate the Moore-Penrose inverse using an iterative Newton-Raphson method in the forward process only, while calculating its theoretical gradients only once in the backward process. To further expand applicability (e.g., dense prediction tasks), an efficient symmetric normalization technique is introduced. Extensive experiments on ImageNet, COCO, and ADE20K show that our SOFT significantly improves the computational efficiency of existing ViT variants. With linear complexity, much longer token sequences are permitted by SOFT, resulting in superior trade-off between accuracy and complexity. Code and models are available at https://github.com/fudan-zvg/SOFT.

Sauradip Nag, Xiatian Zhu, Yi-Zhe Song et al.

Existing Temporal Action Detection (TAD) methods typically take a pre-processing step in converting an input varying-length video into a fixed-length snippet representation sequence, before temporal boundary estimation and action classification. This pre-processing step would temporally downsample the video, reducing the inference resolution and hampering the detection performance in the original temporal resolution. In essence, this is due to a temporal quantization error introduced during the resolution downsampling and recovery. This could negatively impact the TAD performance, but is largely ignored by existing methods. To address this problem, in this work we introduce a novel model-agnostic post-processing method without model redesign and retraining. Specifically, we model the start and end points of action instances with a Gaussian distribution for enabling temporal boundary inference at a sub-snippet level. We further introduce an efficient Taylor-expansion based approximation, dubbed as Gaussian Approximated Post-processing (GAP). Extensive experiments demonstrate that our GAP can consistently improve a wide variety of pre-trained off-the-shelf TAD models on the challenging ActivityNet (+0.2% -0.7% in average mAP) and THUMOS (+0.2% -0.5% in average mAP) benchmarks. Such performance gains are already significant and highly comparable to those achieved by novel model designs. Also, GAP can be integrated with model training for further performance gain. Importantly, GAP enables lower temporal resolutions for more efficient inference, facilitating low-resource applications. The code will be available in https://github.com/sauradip/GAP

Zhe Zhang, Gaochang Wu, Jing Zhang et al.

Unsupervised Domain Adaptive Semantic Segmentation (UDA-SS) aims to transfer the supervision from a labeled source domain to an unlabeled target domain. The majority of existing UDA-SS works typically consider images whilst recent attempts have extended further to tackle videos by modeling the temporal dimension. Although the two lines of research share the major challenges -- overcoming the underlying domain distribution shift, their studies are largely independent, resulting in fragmented insights, a lack of holistic understanding, and missed opportunities for cross-pollination of ideas. This fragmentation prevents the unification of methods, leading to redundant efforts and suboptimal knowledge transfer across image and video domains. Under this observation, we advocate unifying the study of UDA-SS across video and image scenarios, enabling a more comprehensive understanding, synergistic advancements, and efficient knowledge sharing. To that end, we explore the unified UDA-SS from a general data augmentation perspective, serving as a unifying conceptual framework, enabling improved generalization, and potential for cross-pollination of ideas, ultimately contributing to the overall progress and practical impact of this field of research. Specifically, we propose a Quad-directional Mixup (QuadMix) method, characterized by tackling distinct point attributes and feature inconsistencies through four-directional paths for intra- and inter-domain mixing in a feature space. To deal with temporal shifts with videos, we incorporate optical flow-guided feature aggregation across spatial and temporal dimensions for fine-grained domain alignment. Extensive experiments show that our method outperforms the state-of-the-art works by large margins on four challenging UDA-SS benchmarks. Our source code and models will be released at https://github.com/ZHE-SAPI/UDASS.

Anindya Mondal, Sauradip Nag, Joaquin M Prada et al.

Existing action recognition methods are typically actor-specific due to the intrinsic topological and apparent differences among the actors. This requires actor-specific pose estimation (e.g., humans vs. animals), leading to cumbersome model design complexity and high maintenance costs. Moreover, they often focus on learning the visual modality alone and single-label classification whilst neglecting other available information sources (e.g., class name text) and the concurrent occurrence of multiple actions. To overcome these limitations, we propose a new approach called 'actor-agnostic multi-modal multi-label action recognition,' which offers a unified solution for various types of actors, including humans and animals. We further formulate a novel Multi-modal Semantic Query Network (MSQNet) model in a transformer-based object detection framework (e.g., DETR), characterized by leveraging visual and textual modalities to represent the action classes better. The elimination of actor-specific model designs is a key advantage, as it removes the need for actor pose estimation altogether. Extensive experiments on five publicly available benchmarks show that our MSQNet consistently outperforms the prior arts of actor-specific alternatives on human and animal single- and multi-label action recognition tasks by up to 50%. Code is made available at https://github.com/mondalanindya/MSQNet.

Kam Woh Ng, Xiatian Zhu, Jiun Tian Hoe et al.

Unsupervised hashing methods typically aim to preserve the similarity between data points in a feature space by mapping them to binary hash codes. However, these methods often overlook the fact that the similarity between data points in the continuous feature space may not be preserved in the discrete hash code space, due to the limited similarity range of hash codes. The similarity range is bounded by the code length and can lead to a problem known as similarity collapse. That is, the positive and negative pairs of data points become less distinguishable from each other in the hash space. To alleviate this problem, in this paper a novel Similarity Distribution Calibration (SDC) method is introduced. SDC aligns the hash code similarity distribution towards a calibration distribution (e.g., beta distribution) with sufficient spread across the entire similarity range, thus alleviating the similarity collapse problem. Extensive experiments show that our SDC outperforms significantly the state-of-the-art alternatives on coarse category-level and instance-level image retrieval. Code is available at https://github.com/kamwoh/sdc.

Victor Escorcia, Ricardo Guerrero, Xiatian Zhu et al. · cambridge

Learning an egocentric action recognition model from video data is challenging due to distractors (e.g., irrelevant objects) in the background. Further integrating object information into an action model is hence beneficial. Existing methods often leverage a generic object detector to identify and represent the objects in the scene. However, several important issues remain. Object class annotations of good quality for the target domain (dataset) are still required for learning good object representation. Besides, previous methods deeply couple the existing action models and need to retrain them jointly with object representation, leading to costly and inflexible integration. To overcome both limitations, we introduce Self-Supervised Learning Over Sets (SOS), an approach to pre-train a generic Objects In Contact (OIC) representation model from video object regions detected by an off-the-shelf hand-object contact detector. Instead of augmenting object regions individually as in conventional self-supervised learning, we view the action process as a means of natural data transformations with unique spatio-temporal continuity and exploit the inherent relationships among per-video object sets. Extensive experiments on two datasets, EPIC-KITCHENS-100 and EGTEA, show that our OIC significantly boosts the performance of multiple state-of-the-art video classification models.

Xian Shi, Xun Xu, Wanyue Zhang et al. · cambridge

Semantic understanding of 3D point cloud relies on learning models with massively annotated data, which, in many cases, are expensive or difficult to collect. This has led to an emerging research interest in semi-supervised learning (SSL) for 3D point cloud. It is commonly assumed in SSL that the unlabeled data are drawn from the same distribution as that of the labeled ones; This assumption, however, rarely holds true in realistic environments. Blindly using out-of-distribution (OOD) unlabeled data could harm SSL performance. In this work, we propose to selectively utilize unlabeled data through sample weighting, so that only conducive unlabeled data would be prioritized. To estimate the weights, we adopt a bi-level optimization framework which iteratively optimizes a metaobjective on a held-out validation set and a task-objective on a training set. Faced with the instability of efficient bi-level optimizers, we further propose three regularization techniques to enhance the training stability. Extensive experiments on 3D point cloud classification and segmentation tasks verify the effectiveness of our proposed method. We also demonstrate the feasibility of a more efficient training strategy.

Junting Pan, Ziyi Lin, Yuying Ge et al.

Video Question Answering (VideoQA) has been significantly advanced from the scaling of recent Large Language Models (LLMs). The key idea is to convert the visual information into the language feature space so that the capacity of LLMs can be fully exploited. Existing VideoQA methods typically take two paradigms: (1) learning cross-modal alignment, and (2) using an off-the-shelf captioning model to describe the visual data. However, the first design needs costly training on many extra multi-modal data, whilst the second is further limited by limited domain generalization. To address these limitations, a simple yet effective Retrieving-to-Answer (R2A) framework is proposed.Given an input video, R2A first retrieves a set of semantically similar texts from a generic text corpus using a pre-trained multi-modal model (e.g., CLIP). With both the question and the retrieved texts, a LLM (e.g., DeBERTa) can be directly used to yield a desired answer. Without the need for cross-modal fine-tuning, R2A allows for all the key components (e.g., LLM, retrieval model, and text corpus) to plug-and-play. Extensive experiments on several VideoQA benchmarks show that despite with 1.3B parameters and no fine-tuning, our R2A can outperform the 61 times larger Flamingo-80B model even additionally trained on nearly 2.1B multi-modal data.

Lihua Zhou, Mao Ye, Xiatian Zhu et al.

Open-vocabulary object detection with vision-language models (VLMs) such as Grounding DINO suffers from performance degradation under test-time distribution shifts, primarily due to semantic misalignment between text embeddings and shifted visual embeddings of region proposals. While recent test-time adaptive object detection methods for VLM-based either rely on costly backpropagation or bypass semantic misalignment via external memory, none directly and efficiently align text and vision in a training-free manner. To address this, we propose Reward-Guided Semantic Evolution (RGSE), a training-free framework that directly refines the text embeddings at test time. Inspired by evolutionary search, RGSE treats text embedding adaptation as a semantic search process: it perturbs text embeddings as candidate variants, evaluates them via cosine similarity with current and historical high-confidence visual proposals as a reward signal, and fuses them into a refined embedding through reward-weighted averaging. Without any backpropagation, RGSE achieves state-of-the-art performance across multiple detection benchmarks while adding minimal computational overhead. Our code will be open source upon publication.

Hengyuan Ma, Xiatian Zhu, Jianfeng Feng et al.

Score-based generative models (SGMs) have recently emerged as a promising class of generative models. However, a fundamental limitation is that their sampling process is slow due to a need for many (e.g., 2000) iterations of sequential computations. An intuitive acceleration method is to reduce the sampling iterations which however causes severe performance degradation. We assault this problem to the ill-conditioned issues of the Langevin dynamics and reverse diffusion in the sampling process. Under this insight, we propose a novel preconditioned diffusion sampling (PDS) method that leverages matrix preconditioning to alleviate the aforementioned problem. PDS alters the sampling process of a vanilla SGM at marginal extra computation cost and without model retraining. Theoretically, we prove that PDS preserves the output distribution of the SGM, with no risk of inducing systematical bias to the original sampling process. We further theoretically reveal a relation between the parameter of PDS and the sampling iterations, easing the parameter estimation under varying sampling iterations. Extensive experiments on various image datasets with a variety of resolutions and diversity validate that our PDS consistently accelerates off-the-shelf SGMs whilst maintaining the synthesis quality. In particular, PDS can accelerate by up to 28x on more challenging high-resolution (1024x1024) image generation. Compared with the latest generative models (e.g., CLD-SGM and Analytic-DDIM), PDS can achieve the best sampling quality on CIFAR-10 at an FID score of 1.99. Our code is publicly available to foster any further research https://github.com/fudan-zvg/PDS.

Xiao Han, Yukang Cao, Kai Han et al.

Recently, text-guided 3D generative methods have made remarkable advancements in producing high-quality textures and geometry, capitalizing on the proliferation of large vision-language and image diffusion models. However, existing methods still struggle to create high-fidelity 3D head avatars in two aspects: (1) They rely mostly on a pre-trained text-to-image diffusion model whilst missing the necessary 3D awareness and head priors. This makes them prone to inconsistency and geometric distortions in the generated avatars. (2) They fall short in fine-grained editing. This is primarily due to the inherited limitations from the pre-trained 2D image diffusion models, which become more pronounced when it comes to 3D head avatars. In this work, we address these challenges by introducing a versatile coarse-to-fine pipeline dubbed HeadSculpt for crafting (i.e., generating and editing) 3D head avatars from textual prompts. Specifically, we first equip the diffusion model with 3D awareness by leveraging landmark-based control and a learned textual embedding representing the back view appearance of heads, enabling 3D-consistent head avatar generations. We further propose a novel identity-aware editing score distillation strategy to optimize a textured mesh with a high-resolution differentiable rendering technique. This enables identity preservation while following the editing instruction. We showcase HeadSculpt's superior fidelity and editing capabilities through comprehensive experiments and comparisons with existing methods.

Li Zhang, Jiachen Lu, Sixiao Zheng et al.

The emergence of vision transformers (ViTs) in image classification has shifted the methodologies for visual representation learning. In particular, ViTs learn visual representation at full receptive field per layer across all the image patches, in comparison to the increasing receptive fields of CNNs across layers and other alternatives (e.g., large kernels and atrous convolution). In this work, for the first time we explore the global context learning potentials of ViTs for dense visual prediction (e.g., semantic segmentation). Our motivation is that through learning global context at full receptive field layer by layer, ViTs may capture stronger long-range dependency information, critical for dense prediction tasks. We first demonstrate that encoding an image as a sequence of patches, a vanilla ViT without local convolution and resolution reduction can yield stronger visual representation for semantic segmentation. For example, our model, termed as SEgmentation TRansformer (SETR), excels on ADE20K (50.28% mIoU, the first position in the test leaderboard on the day of submission) and performs competitively on Cityscapes. However, the basic ViT architecture falls short in broader dense prediction applications, such as object detection and instance segmentation, due to its lack of a pyramidal structure, high computational demand, and insufficient local context. For tackling general dense visual prediction tasks in a cost-effective manner, we further formulate a family of Hierarchical Local-Global (HLG) Transformers, characterized by local attention within windows and global-attention across windows in a pyramidal architecture. Extensive experiments show that our methods achieve appealing performance on a variety of dense prediction tasks (e.g., object detection and instance segmentation and semantic segmentation) as well as image classification.

Yanqin Jiang, Li Zhang, Zhenwei Miao et al.

3D object detection in autonomous driving aims to reason "what" and "where" the objects of interest present in a 3D world. Following the conventional wisdom of previous 2D object detection, existing methods often adopt the canonical Cartesian coordinate system with perpendicular axis. However, we conjugate that this does not fit the nature of the ego car's perspective, as each onboard camera perceives the world in shape of wedge intrinsic to the imaging geometry with radical (non-perpendicular) axis. Hence, in this paper we advocate the exploitation of the Polar coordinate system and propose a new Polar Transformer (PolarFormer) for more accurate 3D object detection in the bird's-eye-view (BEV) taking as input only multi-camera 2D images. Specifically, we design a cross attention based Polar detection head without restriction to the shape of input structure to deal with irregular Polar grids. For tackling the unconstrained object scale variations along Polar's distance dimension, we further introduce a multi-scalePolar representation learning strategy. As a result, our model can make best use of the Polar representation rasterized via attending to the corresponding image observation in a sequence-to-sequence fashion subject to the geometric constraints. Thorough experiments on the nuScenes dataset demonstrate that our PolarFormer outperforms significantly state-of-the-art 3D object detection alternatives.

Zeyu Yang, Jiaqi Chen, Zhenwei Miao et al.

Existing top-performance 3D object detectors typically rely on the multi-modal fusion strategy. This design is however fundamentally restricted due to overlooking the modality-specific useful information and finally hampering the model performance. To address this limitation, in this work we introduce a novel modality interaction strategy where individual per-modality representations are learned and maintained throughout for enabling their unique characteristics to be exploited during object detection. To realize this proposed strategy, we design a DeepInteraction architecture characterized by a multi-modal representational interaction encoder and a multi-modal predictive interaction decoder. Experiments on the large-scale nuScenes dataset show that our proposed method surpasses all prior arts often by a large margin. Crucially, our method is ranked at the first position at the highly competitive nuScenes object detection leaderboard.

Song Tang, Yunxiang Bai, Wenxin Su et al.

Source-Free Domain Adaptation (SFDA) seeks to adapt a source model, which is pre-trained on a supervised source domain, for a target domain, with only access to unlabeled target training data. Relying on pseudo labeling and/or auxiliary supervision, conventional methods are inevitably error-prone. To mitigate this limitation, in this work we for the first time explore the potentials of off-the-shelf vision-language (ViL) multimodal models (e.g., CLIP) with rich whilst heterogeneous knowledge. We find that directly applying the ViL model to the target domain in a zero-shot fashion is unsatisfactory, as it is not specialized for this particular task but largely generic. To make it task-specific, we propose a novel DIFO++ approach. Specifically, DIFO++ alternates between two steps during adaptation: (i) Customizing the ViL model by maximizing the mutual information with the target model in a prompt learning manner, (ii) Distilling the knowledge of this customized ViL model to the target model, centering on gap region reduction. During progressive knowledge adaptation, we first identify and focus on the gap region, where enclosed features are entangled and class-ambiguous, as it often captures richer task-specific semantics. Reliable pseudo-labels are then generated by fusing predictions from the target and ViL models, supported by a memory mechanism. Finally, gap region reduction is guided by category attention and predictive consistency for semantic alignment, complemented by referenced entropy minimization to suppress uncertainty. Extensive experiments show that DIFO++ significantly outperforms the state-of-the-art alternatives. Our code and data are available at https://github.com/tntek/DIFO-Plus.

Xi Chen, Haosen Yang, Sheng Jin et al.

Open-vocabulary segmentation poses significant challenges, as it requires segmenting and recognizing objects across an open set of categories in unconstrained environments. Building on the success of powerful vision-language (ViL) foundation models, such as CLIP, recent efforts sought to harness their zero-short capabilities to recognize unseen categories. Despite notable performance improvements, these models still encounter the critical issue of generating precise mask proposals for unseen categories and scenarios, resulting in inferior segmentation performance eventually. To address this challenge, we introduce a novel approach, FrozenSeg, designed to integrate spatial knowledge from a localization foundation model (e.g., SAM) and semantic knowledge extracted from a ViL model (e.g., CLIP), in a synergistic framework. Taking the ViL model's visual encoder as the feature backbone, we inject the space-aware feature into the learnable queries and CLIP features within the transformer decoder. In addition, we devise a mask proposal ensemble strategy for further improving the recall rate and mask quality. To fully exploit pre-trained knowledge while minimizing training overhead, we freeze both foundation models, focusing optimization efforts solely on a lightweight transformer decoder for mask proposal generation-the performance bottleneck. Extensive experiments demonstrate that FrozenSeg advances state-of-the-art results across various segmentation benchmarks, trained exclusively on COCO panoptic data, and tested in a zero-shot manner. Code is available at https://github.com/chenxi52/FrozenSeg.

Sauradip Nag, Anran Qi, Xiatian Zhu et al.

Garment pattern design aims to convert a 3D garment to the corresponding 2D panels and their sewing structure. Existing methods rely either on template fitting with heuristics and prior assumptions, or on model learning with complicated shape parameterization. Importantly, both approaches do not allow for personalization of the output garment, which today has increasing demands. To fill this demand, we introduce PersonalTailor: a personalized 2D pattern design method, where the user can input specific constraints or demands (in language or sketch) for personal 2D panel fabrication from 3D point clouds. PersonalTailor first learns a multi-modal panel embeddings based on unsupervised cross-modal association and attentive fusion. It then predicts a binary panel masks individually using a transformer encoder-decoder framework. Extensive experiments show that our PersonalTailor excels on both personalized and standard pattern fabrication tasks.

Yurui Chen, Chun Gu, Junzhe Jiang et al.

Modeling dynamic, large-scale urban scenes is challenging due to their highly intricate geometric structures and unconstrained dynamics in both space and time. Prior methods often employ high-level architectural priors, separating static and dynamic elements, resulting in suboptimal capture of their synergistic interactions. To address this challenge, we present a unified representation model, called Periodic Vibration Gaussian (PVG). PVG builds upon the efficient 3D Gaussian splatting technique, originally designed for static scene representation, by introducing periodic vibration-based temporal dynamics. This innovation enables PVG to elegantly and uniformly represent the characteristics of various objects and elements in dynamic urban scenes. To enhance temporally coherent and large scene representation learning with sparse training data, we introduce a novel temporal smoothing mechanism and a position-aware adaptive control strategy respectively. Extensive experiments on Waymo Open Dataset and KITTI benchmarks demonstrate that PVG surpasses state-of-the-art alternatives in both reconstruction and novel view synthesis for both dynamic and static scenes. Notably, PVG achieves this without relying on manually labeled object bounding boxes or expensive optical flow estimation. Moreover, PVG exhibits 900-fold acceleration in rendering over the best alternative.

Mingxiu Cai, Zhe Zhang, Gaochang Wu et al.

Unsupervised Anomaly Detection (UAD) aims to identify abnormal regions by establishing correspondences between test images and normal templates. Existing methods primarily rely on image reconstruction or template retrieval but face a fundamental challenge: matching between test images and normal templates inevitably introduces noise due to intra-class variations, imperfect correspondences, and limited templates. Observing that Retrieval-Augmented Generation (RAG) leverages retrieved samples directly in the generation process, we reinterpret UAD through this lens and introduce \textbf{RAID}, a retrieval-augmented UAD framework designed for noise-resilient anomaly detection and localization. Unlike standard RAG that enriches context or knowledge, we focus on using retrieved normal samples to guide noise suppression in anomaly map generation. RAID retrieves class-, semantic-, and instance-level representations from a hierarchical vector database, forming a coarse-to-fine pipeline. A matching cost volume correlates the input with retrieved exemplars, followed by a guided Mixture-of-Experts (MoE) network that leverages the retrieved samples to adaptively suppress matching noise and produce fine-grained anomaly maps. RAID achieves state-of-the-art performance across full-shot, few-shot, and multi-dataset settings on MVTec, VisA, MPDD, and BTAD benchmarks. \href{https://github.com/Mingxiu-Cai/RAID}{https://github.com/Mingxiu-Cai/RAID}.

Jiaqi Chen, Jiachen Lu, Xiatian Zhu et al.

We present Generative Semantic Segmentation (GSS), a generative learning approach for semantic segmentation. Uniquely, we cast semantic segmentation as an image-conditioned mask generation problem. This is achieved by replacing the conventional per-pixel discriminative learning with a latent prior learning process. Specifically, we model the variational posterior distribution of latent variables given the segmentation mask. To that end, the segmentation mask is expressed with a special type of image (dubbed as maskige). This posterior distribution allows to generate segmentation masks unconditionally. To achieve semantic segmentation on a given image, we further introduce a conditioning network. It is optimized by minimizing the divergence between the posterior distribution of maskige (i.e., segmentation masks) and the latent prior distribution of input training images. Extensive experiments on standard benchmarks show that our GSS can perform competitively to prior art alternatives in the standard semantic segmentation setting, whilst achieving a new state of the art in the more challenging cross-domain setting.

Huanxin Chen, Pengshuai Yin, Huichou Huang et al.

Predicting typhoon intensity accurately across space and time is crucial for issuing timely disaster warnings and facilitating emergency response. This has vast potential for minimizing life losses and property damages as well as reducing economic and environmental impacts. Leveraging satellite imagery for scenario analysis is effective but also introduces additional challenges due to the complex relations among clouds and the highly dynamic context. Existing deep learning methods in this domain rely on convolutional neural networks (CNNs), which suffer from limited per-layer receptive fields. This limitation hinders their ability to capture long-range dependencies and global contextual knowledge during inference. In response, we introduce a novel approach, namely "Typhoon Intensity Transformer" (Tint), which leverages self-attention mechanisms with global receptive fields per layer. Tint adopts a sequence-to-sequence feature representation learning perspective. It begins by cutting a given satellite image into a sequence of patches and recursively employs self-attention operations to extract both local and global contextual relations between all patch pairs simultaneously, thereby enhancing per-patch feature representation learning. Extensive experiments on a publicly available typhoon benchmark validate the efficacy of Tint in comparison with both state-of-the-art deep learning and conventional meteorological methods. Our code is available at https://github.com/chen-huanxin/Tint.

Kam Woh Ng, Xiatian Zhu, Yi-Zhe Song et al.

Recent text-to-image (T2I) generative models allow for high-quality synthesis following either text instructions or visual examples. Despite their capabilities, these models face limitations in creating new, detailed creatures within specific categories (e.g., virtual dog or bird species), which are valuable in digital asset creation and biodiversity analysis. To bridge this gap, we introduce a novel task, Virtual Creatures Generation: Given a set of unlabeled images of the target concepts (e.g., 200 bird species), we aim to train a T2I model capable of creating new, hybrid concepts within diverse backgrounds and contexts. We propose a new method called DreamCreature, which identifies and extracts the underlying sub-concepts (e.g., body parts of a specific species) in an unsupervised manner. The T2I thus adapts to generate novel concepts (e.g., new bird species) with faithful structures and photorealistic appearance by seamlessly and flexibly composing learned sub-concepts. To enhance sub-concept fidelity and disentanglement, we extend the textual inversion technique by incorporating an additional projector and tailored attention loss regularization. Extensive experiments on two fine-grained image benchmarks demonstrate the superiority of DreamCreature over prior methods in both qualitative and quantitative evaluation. Ultimately, the learned sub-concepts facilitate diverse creative applications, including innovative consumer product designs and nuanced property modifications.

Jiachen Lu, Zheyuan Zhou, Xiatian Zhu et al.

A self-driving perception model aims to extract 3D semantic representations from multiple cameras collectively into the bird's-eye-view (BEV) coordinate frame of the ego car in order to ground downstream planner. Existing perception methods often rely on error-prone depth estimation of the whole scene or learning sparse virtual 3D representations without the target geometry structure, both of which remain limited in performance and/or capability. In this paper, we present a novel end-to-end architecture for ego 3D representation learning from an arbitrary number of unconstrained camera views. Inspired by the ray tracing principle, we design a polarized grid of "imaginary eyes" as the learnable ego 3D representation and formulate the learning process with the adaptive attention mechanism in conjunction with the 3D-to-2D projection. Critically, this formulation allows extracting rich 3D representation from 2D images without any depth supervision, and with the built-in geometry structure consistent w.r.t. BEV. Despite its simplicity and versatility, extensive experiments on standard BEV visual tasks (e.g., camera-based 3D object detection and BEV segmentation) show that our model outperforms all state-of-the-art alternatives significantly, with an extra advantage in computational efficiency from multi-task learning.

Hengyuan Ma, Li Zhang, Xiatian Zhu et al.

Score-based generative models (SGMs) have recently emerged as a promising class of generative models. However, a fundamental limitation is that their inference is very slow due to a need for many (e.g., 2000) iterations of sequential computations. An intuitive acceleration method is to reduce the sampling iterations which however causes severe performance degradation. We investigate this problem by viewing the diffusion sampling process as a Metropolis adjusted Langevin algorithm, which helps reveal the underlying cause to be ill-conditioned curvature. Under this insight, we propose a model-agnostic preconditioned diffusion sampling (PDS) method that leverages matrix preconditioning to alleviate the aforementioned problem. Crucially, PDS is proven theoretically to converge to the original target distribution of a SGM, no need for retraining. Extensive experiments on three image datasets with a variety of resolutions and diversity validate that PDS consistently accelerates off-the-shelf SGMs whilst maintaining the synthesis quality. In particular, PDS can accelerate by up to 29x on more challenging high resolution (1024x1024) image generation.

Hengyuan Ma, Li Zhang, Xiatian Zhu et al.

Score-based generative models (SGMs) have recently emerged as a promising class of generative models. The key idea is to produce high-quality images by recurrently adding Gaussian noises and gradients to a Gaussian sample until converging to the target distribution, a.k.a. the diffusion sampling. To ensure stability of convergence in sampling and generation quality, however, this sequential sampling process has to take a small step size and many sampling iterations (e.g., 2000). Several acceleration methods have been proposed with focus on low-resolution generation. In this work, we consider the acceleration of high-resolution generation with SGMs, a more challenging yet more important problem. We prove theoretically that this slow convergence drawback is primarily due to the ignorance of the target distribution. Further, we introduce a novel Target Distribution Aware Sampling (TDAS) method by leveraging the structural priors in space and frequency domains. Extensive experiments on CIFAR-10, CelebA, LSUN, and FFHQ datasets validate that TDAS can consistently accelerate state-of-the-art SGMs, particularly on more challenging high resolution (1024x1024) image generation tasks by up to 18.4x, whilst largely maintaining the synthesis quality. With fewer sampling iterations, TDAS can still generate good quality images. In contrast, the existing methods degrade drastically or even fails completely

Zijie Pan, Jiachen Lu, Xiatian Zhu et al.

High-resolution 3D object generation remains a challenging task primarily due to the limited availability of comprehensive annotated training data. Recent advancements have aimed to overcome this constraint by harnessing image generative models, pretrained on extensive curated web datasets, using knowledge transfer techniques like Score Distillation Sampling (SDS). Efficiently addressing the requirements of high-resolution rendering often necessitates the adoption of latent representation-based models, such as the Latent Diffusion Model (LDM). In this framework, a significant challenge arises: To compute gradients for individual image pixels, it is necessary to backpropagate gradients from the designated latent space through the frozen components of the image model, such as the VAE encoder used within LDM. However, this gradient propagation pathway has never been optimized, remaining uncontrolled during training. We find that the unregulated gradients adversely affect the 3D model's capacity in acquiring texture-related information from the image generative model, leading to poor quality appearance synthesis. To address this overarching challenge, we propose an innovative operation termed Pixel-wise Gradient Clipping (PGC) designed for seamless integration into existing 3D generative models, thereby enhancing their synthesis quality. Specifically, we control the magnitude of stochastic gradients by clipping the pixel-wise gradients efficiently, while preserving crucial texture-related gradient directions. Despite this simplicity and minimal extra cost, extensive experiments demonstrate the efficacy of our PGC in enhancing the performance of existing 3D generative models for high-resolution object rendering.

Swapnil Bhosale, Sauradip Nag, Diptesh Kanojia et al.

Sound Event Detection (SED) aims to predict the temporal boundaries of all the events of interest and their class labels, given an unconstrained audio sample. Taking either the splitand-classify (i.e., frame-level) strategy or the more principled event-level modeling approach, all existing methods consider the SED problem from the discriminative learning perspective. In this work, we reformulate the SED problem by taking a generative learning perspective. Specifically, we aim to generate sound temporal boundaries from noisy proposals in a denoising diffusion process, conditioned on a target audio sample. During training, our model learns to reverse the noising process by converting noisy latent queries to the groundtruth versions in the elegant Transformer decoder framework. Doing so enables the model generate accurate event boundaries from even noisy queries during inference. Extensive experiments on the Urban-SED and EPIC-Sounds datasets demonstrate that our model significantly outperforms existing alternatives, with 40+% faster convergence in training.

Haosen Yang, Chuofan Ma, Bin Wen et al.

Understanding the semantics of individual regions or patches of unconstrained images, such as open-world object detection, remains a critical yet challenging task in computer vision. Building on the success of powerful image-level vision-language (ViL) foundation models like CLIP, recent efforts have sought to harness their capabilities by either training a contrastive model from scratch with an extensive collection of region-label pairs or aligning the outputs of a detection model with image-level representations of region proposals. Despite notable progress, these approaches are plagued by computationally intensive training requirements, susceptibility to data noise, and deficiency in contextual information. To address these limitations, we explore the synergistic potential of off-the-shelf foundation models, leveraging their respective strengths in localization and semantics. We introduce a novel, generic, and efficient architecture, named RegionSpot, designed to integrate position-aware localization knowledge from a localization foundation model (e.g., SAM) with semantic information from a ViL model (e.g., CLIP). To fully exploit pretrained knowledge while minimizing training overhead, we keep both foundation models frozen, focusing optimization efforts solely on a lightweight attention-based knowledge integration module. Extensive experiments in open-world object recognition show that our RegionSpot achieves significant performance gain over prior alternatives, along with substantial computational savings (e.g., training our model with 3 million data in a single day using 8 V100 GPUs). RegionSpot outperforms GLIP-L by 2.9 in mAP on LVIS val set, with an even larger margin of 13.1 AP for more challenging and rare categories, and a 2.5 AP increase on ODinW. Furthermore, it exceeds GroundingDINO-L by 11.0 AP for rare categories on the LVIS minival set.

Swapnil Bhosale, Abhra Chaudhuri, Alex Lee Robert Williams et al.

The introduction of the MUStARD dataset, and its emotion recognition extension MUStARD++, have identified sarcasm to be a multi-modal phenomenon -- expressed not only in natural language text, but also through manners of speech (like tonality and intonation) and visual cues (facial expression). With this work, we aim to perform a rigorous benchmarking of the MUStARD++ dataset by considering state-of-the-art language, speech, and visual encoders, for fully utilizing the totality of the multi-modal richness that it has to offer, achieving a 2\% improvement in macro-F1 over the existing benchmark. Additionally, to cure the imbalance in the `sarcasm type' category in MUStARD++, we propose an extension, which we call \emph{MUStARD++ Balanced}, benchmarking the same with instances from the extension split across both train and test sets, achieving a further 2.4\% macro-F1 boost. The new clips were taken from a novel source -- the TV show, House MD, which adds to the diversity of the dataset, and were manually annotated by multiple annotators with substantial inter-annotator agreement in terms of Cohen's kappa and Krippendorf's alpha. Our code, extended data, and SOTA benchmark models are made public.

Song Tang, Wenxin Su, Mao Ye et al.

Source-Free Domain Adaptation (SFDA) aims to adapt a source model for a target domain, with only access to unlabeled target training data and the source model pre-trained on a supervised source domain. Relying on pseudo labeling and/or auxiliary supervision, conventional methods are inevitably error-prone. To mitigate this limitation, in this work we for the first time explore the potentials of off-the-shelf vision-language (ViL) multimodal models (e.g.,CLIP) with rich whilst heterogeneous knowledge. We find that directly applying the ViL model to the target domain in a zero-shot fashion is unsatisfactory, as it is not specialized for this particular task but largely generic. To make it task specific, we propose a novel Distilling multimodal Foundation model(DIFO)approach. Specifically, DIFO alternates between two steps during adaptation: (i) Customizing the ViL model by maximizing the mutual information with the target model in a prompt learning manner, (ii) Distilling the knowledge of this customized ViL model to the target model. For more fine-grained and reliable distillation, we further introduce two effective regularization terms, namely most-likely category encouragement and predictive consistency. Extensive experiments show that DIFO significantly outperforms the state-of-the-art alternatives. Code is here

Zhi Qin Tan, Olga Isupova, Gustavo Carneiro et al.

Acquiring fine-grained object detection annotations in unconstrained images is time-consuming, expensive, and prone to noise, especially in crowdsourcing scenarios. Most prior object detection methods assume accurate annotations; A few recent works have studied object detection with noisy crowdsourced annotations, with evaluation on distinct synthetic crowdsourced datasets of varying setups under artificial assumptions. To address these algorithmic limitations and evaluation inconsistency, we first propose a novel Bayesian Detector Combination (BDC) framework to more effectively train object detectors with noisy crowdsourced annotations, with the unique ability of automatically inferring the annotators' label qualities. Unlike previous approaches, BDC is model-agnostic, requires no prior knowledge of the annotators' skill level, and seamlessly integrates with existing object detection models. Due to the scarcity of real-world crowdsourced datasets, we introduce large synthetic datasets by simulating varying crowdsourcing scenarios. This allows consistent evaluation of different models at scale. Extensive experiments on both real and synthetic crowdsourced datasets show that BDC outperforms existing state-of-the-art methods, demonstrating its superiority in leveraging crowdsourced data for object detection. Our code and data are available at https://github.com/zhiqin1998/bdc.

Swapnil Bhosale, Haosen Yang, Diptesh Kanojia et al.

Audio-Visual Segmentation (AVS) aims to precisely outline audible objects in a visual scene at the pixel level. Existing AVS methods require fine-grained annotations of audio-mask pairs in supervised learning fashion. This limits their scalability since it is time consuming and tedious to acquire such cross-modality pixel level labels. To overcome this obstacle, in this work we introduce unsupervised audio-visual segmentation with no need for task-specific data annotations and model training. For tackling this newly proposed problem, we formulate a novel Cross-Modality Semantic Filtering (CMSF) approach to accurately associate the underlying audio-mask pairs by leveraging the off-the-shelf multi-modal foundation models (e.g., detection [1], open-world segmentation [2] and multi-modal alignment [3]). Guiding the proposal generation by either audio or visual cues, we design two training-free variants: AT-GDINO-SAM and OWOD-BIND. Extensive experiments on the AVS-Bench dataset show that our unsupervised approach can perform well in comparison to prior art supervised counterparts across complex scenarios with multiple auditory objects. Particularly, in situations where existing supervised AVS methods struggle with overlapping foreground objects, our models still excel in accurately segmenting overlapped auditory objects. Our code will be publicly released.

Qiqi Liu, Huan Xu, Jingyu Li et al.

Autonomous driving requires reasoning about how the environment evolves and planning actions accordingly. Existing world-model-based approaches typically predict future scenes first and plan afterwards, resulting in open-loop imagination that may drift from the actual decision process. In this paper, we present Uni-World VLA, a unified vision-language-action (VLA) model that tightly interleaves future frame prediction and trajectory planning. Instead of generating a full world rollout before planning, our model alternates between predicting future frames and ego actions step by step, allowing planning decisions to be continuously conditioned on the imagined future observations. This interleaved generation forms a closed-loop interaction between world modeling and control, enabling more adaptive decision-making in dynamic traffic scenarios. In addition, we incorporate monocular depth information into frames to provide stronger geometric cues for world modeling, improving long-horizon scene prediction. Experiments on the NAVSIM benchmark show that our approach achieves competitive closed-loop planning performance while producing high-fidelity future frame predictions. These results demonstrate that tightly coupling world prediction and planning is a promising direction for scalable VLA driving systems.

Wenqing Wang, Da Li, Xiatian Zhu et al.

Fine-tuning vision-language models (VLMs) such as CLIP often leads to catastrophic forgetting of pretrained knowledge. Prior work primarily aims to mitigate forgetting during adaptation; however, forgetting often remains inevitable during this process. We introduce a novel paradigm, continued fine-tuning (CFT), which seeks to recover pretrained knowledge after a zero-shot model has already been adapted. We propose a simple, model-agnostic CFT strategy (named MERGETUNE) guided by linear mode connectivity (LMC), which can be applied post hoc to existing fine-tuned models without requiring architectural changes. Given a fine-tuned model, we continue fine-tuning its trainable parameters (e.g., soft prompts or linear heads) to search for a continued model which has two low-loss paths to the zero-shot (e.g., CLIP) and the fine-tuned (e.g., CoOp) solutions. By exploiting the geometry of the loss landscape, the continued model implicitly merges the two solutions, restoring pretrained knowledge lost in the fine-tuned counterpart. A challenge is that the vanilla LMC constraint requires data replay from the pretraining task. We approximate this constraint for the zero-shot model via a second-order surrogate, eliminating the need for large-scale data replay. Experiments show that MERGETUNE improves the harmonic mean of CoOp by +5.6% on base-novel generalisation without adding parameters. On robust fine-tuning evaluations, the LMC-merged model from MERGETUNE surpasses ensemble baselines with lower inference cost, achieving further gains and state-of-the-art results when ensembled with the zero-shot model. Our code is available at https://github.com/Surrey-UP-Lab/MERGETUNE.