WACV 2026 论文列表

LiDAR-based tracking-by-attention (TBA) frameworks inherently suffer from high false negative errors, leading to a significant performance gap compared to traditional LiDAR-based tracking-by-detection (TBD) methods. This paper introduces SCATR, a novel LiDAR-based TBA model designed to address this fundamental challenge systematically. SCATR leverages recent progress in vision-based tracking and incorporates targeted training strategies specifically adapted for LiDAR. Our work's core innovations are two architecture-agnostic training strategies for TBA methods: Second Chance Assignment and Track Query Dropout. Second Chance Assignment is a novel ground truth assignment that concatenates unassigned track queries to the proposal queries before bipartite matching, giving these track queries a second chance to be assigned to a ground truth object and effectively mitigating the conflict between detection and tracking tasks inherent in tracking-by-attention. Track Query Dropout is a training method that diversifies supervised object query configurations to efficiently train the decoder to handle different track query sets, enhancing robustness to missing or newborn tracks. Experiments on the nuScenes tracking benchmark demonstrate that SCATR achieves state-of-the-art performance among LiDAR-based TBA methods, outperforming previous works by 7.6% AMOTA and successfully bridging the long-standing performance gap between LiDAR-based TBA and TBD methods. Ablation studies further validate the effectiveness and generalization of Second Chance Assignment and Track Query Dropout. Code can be found at the following link: \href https://github.com/TRAILab/SCATR https://github.com/TRAILab/SCATR

Inaccurate freespace detection remains a significant challenge to the safety of autonomous driving. However, we observe that current multisource fusion approaches rely on converting LiDAR point clouds into depth maps, often lose crucial 3D geometric cues. This compromises the spatial consistency of predictions, especially in complex urban scenes. To address this limitation, we propose LiDAR-DHMT (LiDAR-Adaptive Dual-branch Hierarchical Mask Transformer), a novel framework designed for spatial-consistent freespace detection and semantic segmentation. Our key innovation lies in the introduction of a 3D Relative Position Bias module, which effectively captures LiDAR's inherent spatial priors. This is coupled with a Dynamic Bias Attention mechanism that adaptively incorporates the 3D positional cues into the Transformer's attention computation, enhancing spatial coherence. Additionally, we employ a Mask Interaction module and a global-local fusion strategy to jointly model contextual semantics and fine-grained structural details. Extensive experiments conducted on the KITTI Road, KITTI-360, Cityscapes datasets demonstrate that LiDAR-DHMT consistently outperforms existing state-of-the-art methods, achieving a competitive 97.59% F1 score in freespace detection and 69.45% and 84.4% mIoU in semantic segmentation. Our findings suggest that LiDAR-DHMT offers a practical solution for deploying robust freespace perception in complex urban driving environments.

Self-supervised learning (SSL) has become a powerful paradigm for learning from large, unlabeled datasets, particularly in computer vision (CV). However, applying SSL to multispectral remote sensing (RS) images presents unique challenges and opportunities due to the geographical and temporal variability of the data. In this paper, we introduce GeoRank, a novel regularization method for contrastive SSL that improves upon prior techniques by directly optimizing spherical distances to embed geographical relationships into the learned feature space. GeoRank outperforms or matches prior methods that integrate geographical metadata and consistently improves diverse contrastive SSL algorithms (e.g., BYOL, DINO). Beyond this, we present a systematic investigation of key adaptations of contrastive SSL for multispectral RS images, including the effectiveness of data augmentations, the impact of dataset cardinality and image size on performance, and the task dependency of temporal views. Code is available at https://github.com/tomburgert/georank.

Visual compatibility recommendation systems aim to surface compatible items (e.g. pants, shoes) that harmonise with a user-selected product (e.g., shirt). Existing methods struggle in three key aspects: they rely on global CNN representations that overlook fine-grained local cues critical for visual pairing; they force all categories into a single latent space, ignoring the fact that compatibility rules differ across product-type pairs; and they demand costly, expert-annotated outfit labels. We introduce MEDAL(Meta-space Distillation and Alignment ), a self-supervised framework that addresses all three challenges simultaneously. MEDAL (i) employs a local-global augmentation curriculum inside a teacher-student ViT to emphasise patch-level texture and pattern similarities while suppressing confounding global shape cues; (ii) partitions the joint feature manifold into learnable, pair-specific meta-spaces so that, for example, (shirt,pants) and (pants,shoes) relationships are modelled with distinct projection masks; and (iii) replaces manual labels with distantly supervised KD, harvesting pseudo-compatible sets via object detection on web images, thus scaling to millions of real-world examples. We further fuse perceptually uniform LUV colour histograms to capture global colour harmony often missed by pure vision transformers. Extensive experiments on Polyvore disjoint/non-disjoint and a 2M-image in-house dataset show state-of-the-art gains of up to +3.72/+2.7FITB and +9.58R@10 over the strongest baseline, whilst cutting annotation cost to zero. Qualitative studies confirm that MEDAL retrieves stylistically coherent outfits and correctly penalises mismatched colour palettes.

Generative pre-training has significantly advanced natural language understanding. Building upon this success, recent research begins to innovate Large Vision Models (LVM) by leveraging large-scale pre-training on visual sequences, where simultaneous consideration of image token sequences within single images and across a set of images is of key importance. This paper shows that sequential modeling on single images and across multiple images can be efficiently and effectively decoupled. We introduce a two-stage learning pipeline, starting with single-image pre-training, followed by fine-tuning on long image/video sequences. We term this method Large Vision Model Lite (LVM-Lite). Extensive experiments showcase the impressive performance of LVM-Lite across various generative and discriminative benchmarks, comparable to specifically trained models without the need for task-specific training. Importantly, LVM-Lite accelerates training speed substantially up to 2.7xand demonstrates strong scalability.

Variations in environment and sensor (ES) conditions--lighting, ISO, shutter, and aperture--cause domain shifts that degrade visual recognition. While a recent robustness benchmark, ImageNet-ES, shows that these shifts differ from conventional augmentations, collecting physically recaptured data is costly and hard to scale. We present CycleGAN-ES, a per-condition unpaired translation framework that simulates ES-style variations from a small set of real targets. Trained on Tiny-ImageNet and ImageNet-ES domain pairs with as few as 200 images per target domain and minimal tuning, CycleGAN-ES produces a synthetic counterpart, ImageNet-sES (IN-sES). The generated images exhibit high-fidelity ES effects both qualitatively and quantitatively, reproducing characteristic exposure and noise behaviors (e.g., highlight clipping at long exposure and increased high-ISO noise). In benchmark evaluations, augmenting training with ImageNet-sES improves robustness to ES shifts on ImageNet-ES, achieves complementary gains when combined with standard augmentation strategies, and transfers to other corruption domains such as ImageNet-C. The learned translators further transfer to new datsets (e.g., CIFAR-100) without retraining. To the best of our knowledge, this is the first systematic study of ES simulation anchored to real recaptures at ImageNet scale. Our results establish ES simulation as a scalable, practical route to incorporating ES-driven style diversity into training pipelines and lay the groundwork for broader real-world robustness evaluation.

While recent 3D generative models can produce high-quality texture images, they often fail to capture human preferences or meet task-specific requirements. Moreover, a core challenge in the 3D texture generation domain is that most existing approaches rely on repeated calls to 2D text-to-image generative models, which lack an inherent understanding of the 3D structure of the input 3D mesh object. To alleviate these issues, we propose an end-to-end differentiable, reinforcement-learning-free framework that embeds human feedback, expressed as differentiable reward functions, directly into the 3D texture synthesis pipeline. By back-propagating preference signals through both geometric and appearance modules of the proposed framework, our method generates textures that respect the 3D geometry structure and align with desired criteria. To demonstrate its versatility, we introduce three novel geometry-aware reward functions, which offer a more controllable and interpretable pathway for creating high-quality 3D content from natural language. By conducting qualitative, quantitative, and user-preference evaluations against state-of-the-art methods, we demonstrate that our proposed strategy consistently outperforms existing approaches. Our implementation code is publicly available at: https://github.com/AHHHZ975/Differentiable-Texture-Learning

Multimodal emotion recognition in conversation (MERC) requires representations that effectively integrate signals from multiple modalities. These signals include modality-specific cues, information shared across modalities, and interactions that emerge only when modalities are combined. In information-theoretic terms, these correspond to unique, redundant, and synergistic contributions. An ideal representation should leverage all three, yet achieving such balance remains challenging. Recent advances in contrastive learning and augmentation-based methods have made progress, but they often overlook the role of data preparation in preserving these components. In particular, applying augmentations directly to raw inputs or fused embeddings can blur the boundaries between modality-unique and cross-modal signals. To address this challenge, we propose a two-phase framework Divide and Refine (DnR). In the Divide phase, each modality is explicitly decomposed into uniqueness, pairwise redundancy, and synergy. In the Refine phase, tailored objectives enhance the informativeness of these components while maintaining their distinct roles. The refined representations are plug-and-play compatible with diverse multimodal pipelines. Extensive experiments on IEMOCAP and MELD demonstrate consistent improvements across multiple MERC backbones. These results highlight the effectiveness of explicitly dividing, refining, and recombining multimodal representations as a principled strategy for advancing emotion recognition.

Roadside monocular 3D detection requires detecting objects of predefined classes in an RGB frame and predicting their 3D attributes, such as bird's-eye-view (BEV) locations. It has broad applications in traffic control, vehicle-vehicle communication, and vehicle-infrastructure cooperative perception. To address this task, we introduce Promptable 3D Detector (Pro3D), a novel detector design that leverages 2D detections as prompts. We build our Pro3D upon two key insights. First, compared to a typical 3D detector, a 2D detector is "easier" to train due to fewer loss terms and performs significantly better at localizing objects w.r.t 2D metrics. Second, once 2D detections precisely locate objects in the image, a 3D detector can focus on lifting these detections into 3D BEV, especially when fixed camera pose or scene geometry provide an informative prior. To encode and incorporate 2D detections, we explore three methods: (a) concatenating features from both 2D and 3D detectors, (b) attentively fusing 2D and 3D detector features, and (c) encoding properties of predicted 2D bounding boxes x, y, width, height, label and attentively fusing them with the 3D detector feature. Interestingly, the third method significantly outperforms the others, underscoring the effectiveness of 2D detections as prompts that offer precise object targets and allow the 3D detector to focus on lifting them into 3D. Pro3D is adaptable for use with a wide range of 2D and 3D detectors with minimal modifications. Comprehensive experiments demonstrate that our Pro3D significantly enhances existing methods, achieving state-of-the-art results on two contemporary benchmarks.

We introduce Eff-GRot, an approach for efficient and generalizable rotation estimation from RGB images. Given a query image and a set of reference images with known orientations, our method directly predicts the object's rotation in a single forward pass, without requiring object- or category-specific training. At the core of our framework is a transformer that performs a comparison in the latent space, jointly processing rotation-aware representations from multiple references alongside a query. This design enables a favorable balance between accuracy and computational efficiency while remaining simple, scalable, and fully end-to-end. Experimental results show that Eff-GRot offers a promising direction toward more efficient rotation estimation, particularly in latency-sensitive applications.

Delineating farm boundaries through segmentation of satellite images is a fundamental step in many agricultural applications. The task is particularly challenging for smallholder farms, where accurate delineation requires the use of high resolution (HR) imagery which are available only at low revisit frequencies (e.g., annually). To support more frequent (sub-) seasonal monitoring, HR images could be combined as references (ref) with low resolution (LR) images - having higher revisit frequency (e.g., weekly) - using reference-based super resolution (Ref-SR) methods. However, current Ref-SR methods optimize perceptual quality and smooth over crucial features needed for downstream tasks, and are unable to meet the large scale-factor requirements for this task. Further, previous two-step approaches of SR followed by segmentation do not effectively utilize diverse satellite sources as inputs. We address these problems through a new approach, SEED-SR, which uses a combination of conditional latent diffusion models and large-scale multi-spectral, multi-source geo-spatial foundation models. Our key innovation is to bypass the explicit SR task in the pixel space and instead perform SR in a segmentation-aware latent space. This unique approach enables us to generate segmentation maps at an unprecedented 20x scale factor, and rigorous experiments on two large, real datasets demonstrate up to 25.5% and 12.9% relative improvement in instance and semantic segmentation metrics respectively over approaches based on state-of-the-art Ref-SR methods

Recent advancement in multimodal LLMs (MLLMs) has demonstrated their remarkable capability to generate descriptive captions for input videos. However, these models suffer from factual inaccuracies in the generated descriptions, causing severe hallucination issues. While prior works have explored alleviating hallucinations for static images, jointly mitigating visual object and temporal action hallucinations for dynamic videos remains a challenging and unsolved task. To tackle this challenge, we propose a Self-Augmented Contrastive Alignment (SANTA) framework for enabling object and action faithfulness by exempting the spurious correlations and enforcing the emphasis on visual facts. SANTA employs a hallucinative self-augmentation scheme to identify the potential hallucinations that lie in the MLLM and transform the original captions to the contrasted negatives. Furthermore, we develop a tracklet-phrase contrastive alignment to match the regional objects and relation-guided actions with their corresponding visual and temporal phrases. Extensive experiments demonstrate that SANTA outperforms existing methods in alleviating object and action hallucinations, yielding superior performance on the hallucination examination benchmarks.

We introduce LighthouseGS, a practical novel view synthesis framework based on 3D Gaussian Splatting that utilizes simple panorama-style captures from a single mobile device. While convenient, this rotation-dominant motion and narrow baseline make accurate camera pose and 3D point estimation challenging, especially in textureless indoor scenes. To address these challenges, LighthouseGS leverages rough geometric priors, such as mobile device camera poses and monocular depth estimation, and utilizes indoor planar structures. Specifically, we propose a new initialization method called plane scaffold assembly to generate consistent 3D points on these structures, followed by a stable pruning strategy to enhance geometry and optimization stability. Additionally, we present geometric and photometric corrections to resolve inconsistencies from motion drift and auto-exposure in mobile devices. Tested on real and synthetic indoor scenes, LighthouseGS delivers photorealistic rendering, outperforming state-of-the-art methods and enabling applications like panoramic view synthesis and object placement. Project page: https://vision3d-lab.github.io/lighthousegs/

Existing SFDA methods struggle to fully use pre-trained knowledge and often rely on a single model's predictions or handcrafted prompts, limiting robustness under domain shift. Multimodal Large Language Models (MLLMs) offer a promising alternative: they encode rich visual-semantic knowledge and generalize well without task-specific tuning. However, their use in SFDA is hindered by instruction-following failures, inconsistent outputs, and high inference costs. We propose Reliability-based Curriculum Learning (RCL), a novel framework that distills robust supervision from multiple frozen MLLMs into a compact target model. RCL organizes adaptation as a three-stage curriculum that progressively incorporates pseudo-labels based on inter-model agreement and model confidence, enabling stable and noise-aware training. Our approach achieves state-of-the-art performance on standard SFDA datasets, Office-Home, DomainNet-126, and VisDA-C, outperforming zero-shot MLLMs, their ensembles, all without accessing source data or tuning foundation models. Code is available at https://github.com/Dong-Jie-Chen/RCL.

Obtaining the precise 3D motion of a table tennis ball from standard monocular videos is a challenging problem, as existing methods trained on synthetic data struggle to generalize to the noisy, imperfect ball and table detections of the real world. This is primarily due to the inherent lack of 3D ground truth trajectories and spin annotations for real-world video. To overcome this, we propose a novel two-stage pipeline that divides the problem into a front-end perception task and a back-end 2D-to-3D uplifting task. This separation allows us to train the front-end components with abundant 2D supervision from our newly created TTHQ dataset, while the back-end uplifting network is trained exclusively on physically-correct synthetic data. We specifically re-engineer the uplifting model to be robust to common real-world artifacts, such as missing detections and varying frame rates. By integrating a ball detector and a table keypoint detector, our approach transforms a proof-of-concept uplifting method into a practical, robust, and high-performing end-to-end application for 3D table tennis trajectory and spin analysis.

Benefiting from the inductive biases learned from large-scale datasets, open-vocabulary semantic segmentation (OVSS) leverages the power of vision-language models, such as CLIP, to achieve remarkable progress without requiring task-specific training. However, due to CLIP's pretraining nature on image-text pairs, it tends to focus on global semantic alignment, resulting in suboptimal performance when associating fine-grained visual regions with text. This leads to noisy and inconsistent predictions, particularly in local areas. We attribute this to a dispersed bias stemming from its contrastive training paradigm, which is difficult to alleviate using CLIP features alone. To address this, we propose a structure-aware feature rectification approach that incorporates instance-specific priors derived directly from the image. Specifically, we construct a region adjacency graph (RAG) based on low-level features (e.g. colour and texture) to capture local structural relationships and use it to refine CLIP features by enhancing local discrimination. Extensive experiments show that our method effectively suppresses segmentation noise, improves region-level consistency, and achieves strong performance on multiple open-vocabulary segmentation benchmarks.

Vision Language Models (VLMs) have rapidly advanced and show strong promise for text-based person search (TBPS), a task that requires capturing fine-grained relationships between images and text to distinguish individuals. Previous methods address these challenges through local alignment, yet they are often prone to shortcut learning and spurious correlations, yielding misalignment. Moreover, injecting prior knowledge can distort intra-modality structure. Motivated by our finding that encoder attention surfaces spatially precise evidence from the earliest training epochs, and to alleviate these issues, we introduce ITSELF, an attention-guided framework for implicit local alignment. At its core, Guided Representation with Attentive Bank (GRAB) converts the model's own attention into an Attentive Bank of high-saliency tokens and applies local objectives on this bank, learning fine-grained correspondences without extra supervision. To make the selection reliable and non-redundant, we introduce Multi-Layer Attention for Robust Selection (MARS), which aggregates attention across layers and performs diversity-aware top-k selection; and Adaptive Token Scheduler (ATS), which schedules the retention budget from coarse to fine over training, preserving context early while progressively focusing on discriminative details. Extensive experiments on three widely used TBPS benchmarks show state-of-the-art performance and strong cross-dataset generalization, confirming the effectiveness and robustness of our approach without additional prior supervision. Our project is publicly available at https://trhuuloc.github.io/itself

Video Scene Graph Generation (VidSGG) aims to represent dynamic visual content by detecting objects and modeling their temporal interactions as structured graphs. Prior studies typically target either coarse-grained box-level or fine-grained panoptic pixel-level VidSGG, often requiring task-specific architectures and multi-stage training pipelines. In this paper, we present UNO (UNified Object-centric VidSGG), a single-stage, unified framework that jointly addresses both tasks within an end-to-end architecture. UNO is designed to minimize task-specific modifications and maximize parameter sharing, enabling generalization across different levels of visual granularity. The core of UNO is an extended slot attention mechanism that decomposes visual features into object and relation slots. To ensure robust temporal modeling, we introduce object temporal consistency learning, which enforces consistent object representations across frames without relying on explicit tracking modules. Additionally, a dynamic triplet prediction module links relation slots to corresponding object pairs, capturing evolving interactions over time. We evaluate UNO on standard box-level and pixel-level VidSGG benchmarks. Results demonstrate that UNO not only achieves competitive performance across both tasks but also offers improved efficiency through a unified, object-centric design.

Open-vocabulary 3D segmentation allows exploration of 3D environments using unrestricted natural language queries. Current approaches to open-vocabulary 3D instance segmentation largely concentrate on recognizing object-level instances but face difficulties when dealing with more fine-grained elements of a scene, such as object parts. Some previous work constructs hierarchical open-vocabulary 3D scene representations by geometric over-segmentation, which can't identify parts with similar geometry. In this work, we introduce PS3, an approach to generate 3D part proposals from multi-view 2D masks. PS3 outperforms baselines that rely on geometric over-segmentation in scene-scale open-vocabulary 3D part segmentation.