Inferring object 3D position and orientation from a single RGB camera image is a foundational computer vision task with many important applications. Traditionally, 3D object detection methods are trained in a fully-supervised setup, requiring LiDAR and vast amounts of human annotations, which are laborious, costly, and do not scale well with the ever-increasing amounts of data being captured. To alleviate the issue, we propose a method to train a 3D object detector from a single RGB camera without domain-specific human annotations, making orders of magnitude more data available for training. The method is approximately 500 times faster than previous work, it is able to disentangle object movement source between subsequent frames and it compensates camera focal length differences to aggregate multiple datasets. The method is evaluated on three public datasets, where despite using no domain-specific human labels, it outperforms prior work by a significant margin. It also shows its versatility as a pre-training tool for fully-supervised training and shows that combining pseudo-labels from multiple datasets can achieve comparable accuracy to using human labels from a single dataset. The source code and model are available at https://github.com/jskvrna/MonoSOWAv2 .
Generalized zero-shot learning (GZSL) addresses the challenging task of recognizing both seen and unseen classes by leveraging shared semantic knowledge. A core challenge in this domain is achieving robust visual-semantic alignment to transfer knowledge from seen classes to novel classes. Current state-of-the-art methods typically fine-tune large-scale visual backbones on scarce training data. However, this approach frequently leads to severe overfitting to seen classes, which significantly degrades performance on novel categories. To mitigate this issue, we propose the Dual Adaptive Visual-Semantic Prompt Collaboration Network (VSPCN+), a novel framework that utilizes prompt-tuning for effective feature adaptation. Our method introduces a dual-prompt mechanism comprising both visual and semantic prompts. The semantic prompts guide the visual encoder to learn visual features that are more semantically consistent with class attributes, while the visual prompts steer the semantic encoder to generate semantic representations that are more visually grounded. This collaborative process enhances the overall visual-semantic consistency. A key innovation of our work is the dynamic generation of instance-adaptive prompts, which contrasts with existing prompt-learning methods that rely on static, global prompts. By tailoring prompts to individual instances, our approach enhances the model’s robustness and generalization capabilities across diverse visual inputs. This collaborative adaptation, guided by our dual-prompt mechanism, allows the visual and semantic encoders to produce consistent representations for effective visual-semantic alignment. Extensive experiments on standard GZSL benchmarks demonstrate that our proposed VSPCN+ performs favorably against several state-of-the-art methods.
Monocular Video-to-4D generation faces the fundamental challenge of inferring plausible 3D geometry and motion from limited single-view inputs. We present MotionStruct4D, a novel approach that discovers and exploits the underlying motion structure of 3D Gaussian Splatting for high-quality video-to-4D generation. Our key insight is that real-world motion can be effectively decomposed into coarse rigid transformations that capture principal movements, complemented by detailed non-rigid deformations that account for fine-grained details. MotionStruct4D introduces: (1) a self-supervised motion structure discovery module that identifies quasi-rigid parts by preserving spatiotemporal relationships without explicit 3D supervision, and (2) a weighted dense-to-sparse optimization architecture that transitions from dense per-Gaussian deformation to sparse control points, effectively integrating rigid and non-rigid motion components through adaptive weighted fusion. This design addresses the parametric imbalance between rigid and non-rigid motions and effectively models 3D movements across different kinematic patterns. To evaluate our approach on challenging scenarios, we curate a comprehensive benchmark dataset featuring substantial object displacement and diverse articulated motion patterns. Experimental results demonstrate MotionStruct4D’s superior performance in motion fidelity and novel viewpoint synthesis quality, while also providing interpretable motion structure decomposition that reveals meaningful quasi-rigid part segmentation.
Multi-view human reconstruction has been extensively studied under simplified settings, yet scaling these methods to robust and efficient multi-person reconstruction in unconstrained environments requires a more general and scalable modeling paradigm. Existing bottom-up methods often rely on accurate camera calibration and explicit cross-view matching, and therefore struggle in multi-person scenarios with severe occlusions and ambiguities. We attribute these limitations to the lack of an explicit and view-agnostic 3D representation for jointly reasoning about human structure and multi-view consistency. Based on this observation, we propose a new top-down paradigm that maintains a unified, instance-centric human-aware 3D space, enabling simultaneous camera calibration, cross-view association, and human reconstruction via cross-modal contrastive learning. Specifically, observations from multiple views are lifted and fused directly into this shared 3D space, where geometric structure, visual appearance, and human-centric semantic cues are jointly encoded at the instance level. To enhance the discriminability and consistency of the 3D representation, we introduce a spatial contrastive learning strategy that aligns 3D features corresponding to the same human instance across different views and modalities, while separating features from different instances. This design allows correspondence reasoning, semantic aggregation, and instance discrimination to be performed natively in the 3D space, which enforces cross-view consistency and improves robustness under severe occlusions. Based on the learned instance-aware 3D representation, we recover structured human body models in a feed-forward manner by regressing SMPL parameters from instance-level 3D human tokens. Extensive experiments demonstrate that adopting a unified human-aware 3D feature space as the core representation leads to robust, accurate, and efficient multi-view human reconstruction in challenging real-world scenarios. The code and data will be made publicly available.
End-to-end driving models achieve strong planning performance with large-scale data but remain constrained by training-data bias and often struggle in complex and challenging scenarios. In contrast, Large Vision-Language Models (LVLMs) exhibit promising scene understanding and reasoning abilities. The path forward lies in merging the strengths of both approaches. Prior attempts using LVLMs to predict trajectories or control signals yield suboptimal results, as LVLMs are not well-suited for precise numerical outputs. We introduce Senna, an autonomous driving framework combining an LVLM (Senna-VLM) with an end-to-end model (Senna-E2E). Senna decouples high-level planning from low-level prediction. Senna-VLM produces natural language planning decisions, and Senna-E2E converts them into precise trajectories. Senna-VLM utilizes a multi-image encoding mechanism and multi-view prompts for efficient scene reasoning. Besides, we introduce planning-oriented QAs and a three-stage training strategy, which together enhance Senna-VLM’s scene understanding and planning capabilities. Comprehensive open-loop and closed-loop experiments show that Senna achieves state-of-the-art planning performance. Notably, with pre-training on a large-scale dataset DriveX and fine-tuning on nuScenes, Senna significantly reduces planning error by 35.59 https://github.com/hustvl/Senna to facilitate future research.
Gait recognition has emerged as a promising biometric for long-distance, low-resolution, and non-cooperative scenarios. Recent deep models achieve impressive accuracy from silhouette sequences, yet their robustness and security remain far from fully understood. In this paper, we study a particularly challenging and practically relevant threat: query-free black-box universal adversarial attacks against silhouette-based gait recognition. We adopt a strict query-free black-box setting, where the attacker has no access to the target model and cannot query the deployed system for any feedback. The goal is to learn a universal perturbation that is shared across subjects, walking conditions, and sequences, and that transfers to unknown models and datasets while remaining visually imperceptible on silhouettes. To this end, we propose GaitUAA, a universal attack framework that learns a single-frame perturbation template on surrogate models and then broadcasts it to entire silhouette sequences. Our design combines a retrieval-oriented attack objective with an edge-aware parameterization: the perturbation magnitude is constrained by an upper bound, and an edge threshold restricts perturbations to silhouette boundaries. Extensive experiments on multiple benchmarks and architectures show that GaitUAA consistently causes substantial performance degradation under cross-dataset, cross-model, and cross-dataset-and-model protocols, driving the rank-1 accuracy of several state-of-the-art models (e.g., DeepGaitV2-2D/3D/P3D, SwinGait) down to single-digit levels in the most challenging settings. These results underscore the urgent need for architecture-aware defenses and robustness-oriented design in future gait recognition systems.
Vision-language models pretrained on large-scale datasets have become the foundation of cross-modal understanding. However, they often overemphasize salient objects and underperform in compositional reasoning that additionally requires fine-grained perception of attributes and inter-object relations. Although compositional data-driven approaches have shown promise in addressing this challenge, low-quality captions undermine robust alignment and consequently hinder fine-grained perception. Specifically, two fundamental issues arise: (1) low-density positive captions restrict the global representation capability of image encoders, causing local alignment. (2) false-negative captions introduce noise into training, resulting in fragile alignment. In response, we propose the Robust and Fine-grained training framework for CLIP-based vision-language models (RoFLIP), enhancing both the robustness and granularity of vision-language alignment. To address the first challenge, RoFLIP employs the Densified Ranking Loss (DRL), which promotes global visual representations by contrasting dense and sparse captions. For the second challenge, RoFLIP incorporates Denoised Cross-modal and Text-modal Contrastive Losses (DCCL DTCL), reducing the impact of noise while effectively leveraging negative captions. In addition, we introduce RoFLIP’s Auxiliary dataset (RoFA), comprising 225K high-quality captions that are fluent and image-related to support training. Quantitative evaluations demonstrate that RoFLIP-CLIP-B/32 achieves an average gain of 7.41% across 5 compositional reasoning benchmarks while maintaining considerable performance on 23 downstream benchmarks. Furthermore, replacing the corresponding modules in LLaVA-v1.5 and Stable Diffusion v1.5 with RoFLIP-CLIP-L/14 also brings performance gains on VQA and image generation tasks, respectively. These results underscore RoFLIP’s compositional reasoning and generalization abilities.
Recent advances in deep learning have enabled remarkable success across computer vision tasks using massive datasets, yet the persistent challenge of domain shift-where training and testing distributions differ-continues to degrade model performance. While Domain Generalization techniques aim to learn invariant representations across multiple domains, their reliance on diverse training domains limits applicability in settings where multiple source domains are not available. This survey examines Single Domain Generalization (SDG) approaches that leverage targeted inductive biases to enhance out-of-domain (OOD) performance using a single source domain, addressing key questions regarding their effectiveness across various domain shifts, different model architectures such as CNNs and Vision transformers, and the impact of shape bias for OOD generalization. These questions are crucial as they directly inform the development of robust models that can reliably operate under real-world conditions, guiding the research community in overcoming limitations of current techniques. Our findings indicate that although SDG methods improve performance under highly domain distinguishable stylized shifts, they exhibit limited robustness to background, correlation and corruption shifts on larger datasets, and increased shape bias does not consistently yield enhanced OOD performance, thereby underscoring the need for further research into developing more resilient and generalizable models. For reproducibility, the implementation and code used in our benchmark are available on our official https://github.com/Kavindya98/SDG-Benchmark GitHub repository.
Point cloud data are ubiquitous in 3D vision applications, yet they often suffer from noise, sparsity, and incompleteness, which significantly limit the performance of downstream tasks. To address these challenges, we propose LatentFlow3D, a unified framework for point cloud denoising, upsampling, and completion through latent space modeling with invertible neural networks. For denoising, we disentangle noise from clean geometry by projecting them onto separate channels in latent space, enabling the recovery of intrinsic clean latent codes. For upsampling, we formulate the task as interpolation in latent space, where adaptive weights are learned from local geometric context. For completion, we exploit geometric symmetry by applying affine transformations in latent space to infer the latent codes of missing regions, which are then fused with partial inputs to reconstruct complete shapes. To establish a bijective and expressive mapping between point clouds and latent representations, we employ invertible residual networks, augmented with a feature injection module that integrates rich local geometric features. Extensive experiments on widely used benchmarks demonstrate that LatentFlow3D achieves SOTA performance across all three tasks.
Visual Autoregressive (VAR) modeling has emerged as a compelling paradigm for image generation, offering superior scalability compared to traditional raster-scan approaches. However, equipping VAR with fine-grained spatial control remains challenging due to its unique hierarchical, next-scale prediction nature. Existing attempts often rely on disruptive injection mechanisms or heavy adapters that compromise generative priors and computational efficiency. In this work, we present SCALAR, a novel framework designed to align structural guidance with the coarse-to-fine generation process of VAR. We propose a Scale-wise Conditional Decoding mechanism, which projects semantic signals from a frozen vision encoder into scale-specific layers of the VAR backbone. To address the parameter explosion in high-resolution text-to-image (T2I) tasks, we further advance this design to SCALAR++. By introducing Layer-Shared Scale-wise Conditioning supplemented with lightweight LoRA modulation, SCALAR++ substantially reduces parameter and memory overhead via weight sharing, while maintaining precise control through layer-specific adaptation. Additionally, we introduce a Unified Control Alignment strategy (SCALAR-Uni) to handle diverse control modalities within a single projection space. Extensive experiments on ImageNet and MultiGen-20 M demonstrate that SCALAR and SCALAR++ achieve state-of-the-art performance, outperforming strong baselines like ControlNet++ and ControlAR in both generation fidelity and condition consistency. All the code will be released at https://github.com/AMAP-ML/SCALAR .
Fine-grained open-vocabulary detection (FG-OVD) challenges models to detect objects described by complex, compositional language. The prevalent strategy of encoding descriptive phrases into a single, holistic vector creates significant semantic ambiguity, hindering true compositional understanding. This ambiguity manifests as two opposing failures: attribute over-representation, where a visually dominant attribute leads to mislocalization, and attribute under-representation, where the model fails to enforce a conjunctive match across all specified attributes, leading to false positives. In this paper, we propose GUIDED++, a framework that systematically resolves both issues through a dual strategy. First, we employ architectural decoupling to separate coarse-grained localization from fine-grained discrimination, effectively mitigating attribute over-representation. Second, we introduce a novel conjunctive multi-attribute verification mechanism to explicitly combat attribute under-representation. This mechanism decomposes a complex query into its atomic attribute conditions, computes a distinct similarity score for each, and uses a minimal aggregation function to enforce a strict logical ‘AND’, ensuring an object satisfies all required attributes. Extensive experiments show that GUIDED++ significantly outperforms existing methods and establishes a new state-of-the-art on challenging FG-OVD benchmarks, demonstrating a more robust approach to compositional visual reasoning.
Recent years have witnessed the widespread adoption of multimodal applications across various real-world scenarios. However, these applications often face significant challenges due to distribution shifts across different domains, which can severely impact model performance and reliability. While domain generalization has been extensively studied for unimodal scenarios, its multimodal counterpart remains underexplored, presenting unique challenges in addressing both modality heterogeneity and domain shifts simultaneously. We argue that an effective multimodal domain generalization framework should not only learn domain invariance but also balance the invariance and specificity of modalities. To bridge this gap, we introduce Adaptive Dual-Objective Feature Learning for Multimodal Domain Generalization (ADMMDG), a framework that separates feature learning into two complementary components: features that capture shared patterns with both modality and domain invariance, and modality specific features that preserve the unique characteristics of each modality while maintaining domain invariance. To achieve this, ADMMDG employs two distinct contrastive learning strategies that focus on aligning features across different domains and modalities. Both strategies are equipped with a dynamic weighting mechanism that assigns higher weights to modal-domain feature pairs exhibiting greater modality or domain discrepancies, thereby effectively capturing the respective feature learning objectives. Extensive experiments on multiple benchmark datasets, utilizing diverse modality combinations, demonstrate the effectiveness of ADMMDG in both multi-source and single-source multimodal domain generalization tasks, as well as in the challenging multimodal open-set domain generalization scenario. Our codes are available at https://github.com/lihongzhao99/ADMMDG .
Traditional video restoration methods are typically designed for a single, known degradation type, while real-world videos often experience time-varying and unknown degradations. Although more general restoration methods have been proposed, they largely overlook the time-varying nature of real-world degradations. In this work, we focus on time-varying unknown degradations, which introduce frame-wise distortions and pose two significant challenges to video restoration. First, the unknown nature of degradations expands the input degradation space and introduces significant uncertainty into the restoration process. Second, time-varying degradations cause inconsistent distortions across frames, hindering the effective aggregation of temporal information from neighboring frames. To address these two challenges, we propose FaverNet, a Frequency-guided All-in-one VidEo Restoration Network, which incorporates a Frequency-discriminative Conditioning Mechanism (FCM) and a Prompt-guided Alignment Mechanism (PAM). Specifically, FCM enhances the degradation-awareness of the restoration process by conditioning the model with degradation cues extracted from frequency domain, where different degradation types are more separable than in the spatial domain. PAM further leverages these degradation cues to assist alignment between neighboring frames, enabling temporally coherent and degradation-robust information aggregation. Furthermore, we build two benchmarks covering both classic low-level distortions (e.g., noise, blur) and complex weather-related degradations (e.g., haze, rain). Experiments show that FaverNet offers an encouraging solution for practical video restoration and achieves superior performance.
Long-tailed visual recognition is fundamentally constrained by the geometric sparsity of tail-class feature manifolds, where data scarcity leads to vacuous representation spaces and collapsed decision boundaries. While existing discriminative approaches such as re-balancing and decoupling mitigate bias, they operate passively on observed samples and fail to replenish the missing semantic information in the tail. Drawing inspiration from the brain’s complementary learning systems, specifically the interplay between predictive coding during wakefulness and generative consolidation during sleep, we propose ManiDream, a bio-inspired framework for active feature manifold completion. The framework operates in a rhythmic two-phase cycle. First, in the Online Perception phase, we introduce an Attentive Predictive Estimator (APE) that detects semantic surprise by measuring attention-weighted reconstruction errors. This mechanism enables the model to precisely pinpoint and store hard examples with high epistemic uncertainty. Second, in the Offline Consolidation phase, we devise a Prototype-Anchored Dreaming (PAD) mechanism. PAD employs a residual generative model to actively synthesize high-fidelity pseudo-features within the geometric vicinity of real prototypes, thereby filling the manifold voids without semantic drift. Crucially, to address the feature energy decay inherent in generative augmentation, we introduce a physics-inspired Metacognitive Norm Rectification strategy that preserves feature-norm consistency and maintains effective gradient magnitudes during classifier optimization. Extensive experiments across five benchmarks, including CIFAR-100-LT, CIFAR-10-LT, ImageNet-LT, iNaturalist 2018, and Places-LT, demonstrate that ManiDream achieves competitive or state-of-the-art performance on standard long-tailed recognition protocols and improves robustness across most distribution-shift settings, supporting the effectiveness of active manifold completion.
High-quality digital pathological images face challenges such as high acquisition costs and difficulties in storage and transmission owing to their considerably large size. Several studies based on single-image super-resolution have been conducted to address these challenges; however, the methods employed in these studies generally exhibit excessive smoothness and optimization instability, failing to generate high-fidelity images. This study presents consistency-restricted pathological image reconstruction (CorPiR), an effective two-stage framework for super-resolution pathological image reconstruction. CorPiR uses a lightweight regression-based method in the first stage to quickly reconstruct low-resolution inputs. In the second stage, a consistency-restricted diffusion (CorD) model is proposed to reconstruct image details to generate high-fidelity pathological images. By introducing a cross-attention mechanism-based module, consistency constrainer (CConstrainer), the CorD model enhances the consistency between generated and input images without compromising visual fidelity. Furthermore, we designed a consistency evaluator module to extract consistency prior information from a low-resolution input image, guiding CConstrainer to focus on reconstructing regions that lack consistency. Extensive experiments on a pan-cancer dataset, with ten cancer types at four different magnification levels, verify the effectiveness of the CorPiR framework, demonstrating superior performance compared to existing state-of-the-art methods qualitatively and quantitatively. In addition, evaluations performed by pathologists and the results from handcrafted texture feature experiments demonstrated that CorPiR could generate images with both high visual fidelity and consistency, providing valuable support for clinical diagnosis and downstream computational tasks. The CorPiR code is publicly available at https://github.com/ZHT150798/CorPiR .
In the past decade, image restoration has witnessed a remarkable performance leap. However, existing methods still face two critical limitations: i) many methods train distortion-specific models, which struggle to address mixed distortions; ii) most methods employ reconstruction loss during training, with little consideration for the intended use of the restored images (e.g., human viewing, detection, or segmentation), leading to a mismatch between training objectives and application goals. In this paper, we propose a scalable framework for image restoration and enhancement built upon the reinforcement learning (RL) paradigm. At the core of our framework, the RL agent is capable of accurately perceiving distortion contexts and dynamically adjusting diverse restoration operators, which ensures our framework to effectively handle mixed-distortions, and generalize well to new tasks with minor additional cost. Furthermore, RL relaxes the differentiability constraints on optimization objectives, which enables our framework to directly optimize the restoration process toward application goals, thereby bridging the gap between training objectives and downstream demands. Extensive objective and subjective evaluations demonstrate that our method not only generates visually pleasing results for human perception but also significantly improves the performance of machine vision tasks such as detection and segmentation. We further validate the framework’s scalability through generalization evaluations on unseen distortions and multi-objective optimization experiments. The results confirm the model’s superior few-shot and zero-shot capabilities compared to existing methods, as well as its flexibility in addressing multiple competing objectives. Finally, comprehensive ablation studies and hyperparameter analyses provide valuable insights into the design choices and behavior of our framework.
Recent advances in 3D Gaussian Splatting (3DGS) have enabled feed-forward, on-the-fly reconstruction of sequential input views. However, existing methods often predict per-pixel Gaussians and combine Gaussians from all views as the scene representation, leading to substantial redundancies and geometric inconsistencies in long-duration video sequences. To address this, we propose SaLon3R, a novel framework for Structure-aware, Long-term 3DGS Reconstruction. Our method eliminates redundancy by introducing compact anchor primitives as a replacement for per-pixel Gaussians. These primitives are derived through a differentiable, saliency-aware Gaussian quantization process designed to preserve fidelity while ensuring a compact representation. Specifically, a foundational 3D reconstruction model is employed to predict a saliency map encoding regional geometric complexity. Guided by this saliency map, we compress redundant Gaussian primitives into compact anchors by prioritizing high-complexity regions. Furthermore, we introduce a 3D Point Transformer to overcome geometric inconsistencies caused by long-term accumulative errors. It refines attributes and saliency of the anchor primitives leveraging the learned spatial structural priors in 3D space. Without known camera parameters or test-time optimization, our approach effectively prunes the redundant 3DGS and resolves artifacts in a single feed-forward pass. Experiments on multiple datasets demonstrate our approach outperform state-of-the-arts on both novel view synthesis and depth estimation, while exhibiting superior efficiency ( > 10 FPS), robustness, and generalization ability for long-term feed-forward Gaussian reconstruction.
Reconstructing multi-human body meshes from a single monocular image is a crucial yet challenging problem in computer vision. This problem requires not only generating individual body mesh models for each person but also estimating the relative 3D positions among subjects to produce a coherent scene representation. In this work, we introduce MUG (Multi-hUman Graph network), which employs a single graph neural network to construct coherent multi-human meshes using only 2D pose data as input. Our approach demonstrates that purely pose-based methods can effectively perform simultaneous depth reasoning and multi-human mesh generation. Existing image-based methods typically rely on lab-collected training datasets with accurate 3D labels; however, these datasets often introduce an image domain gap when applied to in-the-wild testing data or art images due to differences in appearance and context. In contrast, MUG leverages the geometric consistency of 2D poses across diverse datasets, mitigating domain discrepancies. The MUG network operates in three primary phases. Initially, to model the multi-human environment, it processes multi-human 2D poses and constructs a novel heterogeneous graph. This graph connects nodes both across different people and within individuals to capture inter-human interactions and accurately represent body geometry, including skeletal and mesh structures. Subsequently, it employs a dual-branch graph neural network: one branch predicts inter-human depth relations, while the other predicts the root-joint-relative mesh coordinates. Finally, the complete multi-human 3D meshes are constructed by combining the outputs from both branches. Despite the simplicity of using only 2D pose inputs and streamlined network architecture, MUG outperforms existing multi-human mesh estimation methods. This superiority is consistently observed across various datasets from diverse domains, such as RH, MuPoTS-3D, and 3DPW. Qualitative results show that MUG can effectively handle art images, through-wall scenarios, and poor lighting conditions when incorporating advanced 2D pose networks. Both qualitative and quantitative evaluations highlight MUG’s remarkable generalization ability in open-world scenarios.
Reliable 3D reconstruction is a prerequisite for robotics, embodied AI, and immersive AR/VR applications; however, real-world observations frequently depart from clean imaging assumptions due to illumination changes, participating media, occlusions, and blur that break multi-view consistency and destabilize pose estimation, which leaves a gap between performance on curated benchmarks and behavior in practical deployments. To address this gap, we introduce RealX3D, a real capture benchmark for multi-view restoration and reconstruction under real-world degradations, organized into four families spanning nine controlled settings that include motion and defocus blur, low-light, view-varying exposure, smoke, dynamic occlusion, and reflection. RealX3D is collected using a unified acquisition protocol that enables recapturing the same camera trajectories to obtain pixel-aligned low-quality and reference ground-truth image pairs. Each scene also provides per-view RAW measurements to preserve high dynamic range linear sensor signals. To support geometry-grounded evaluation beyond image photometric fidelity, we capture dense laser scan geometry for every scene and derive world-scale measures such as point clouds, meshes, and metric depth, allowing comprehensive assessment of pose, depth, and surface reconstruction alongside photometric restoration quality. The benchmark contains 55 scenes recorded at high resolution with diverse real-world degradation patterns. We benchmark a broad set of optimization-based and feed-forward methods using both image metrics and geometry metrics, and the results reveal substantial robustness gaps across degradations in adverse conditions. Overall, RealX3D provides a rigorous benchmark that moves beyond synthetic data and establishes a standardized foundation for developing degradation-robust 3D reconstruction systems.
Large Vision-Language Models (LVLMs) demonstrate strong capabilities but remain susceptible to hallucinations. To address this limitation, we propose reflective instruction tuning, which explicitly trains models to reflect. Rather than producing only a final answer, the model is supervised to generate reflective rationales that justify the correct prediction and explain why plausible alternatives are incorrect. To remain effective for advanced LVLMs deployed in diverse real-world scenarios, reflective supervision must be more fine-grained and broader in domains. We therefore propose REVERIE+ (ReflEctiVE RatIonalE), an extension of REVERIE substantially expanded in domain diversity, task complexity, and annotation richness, tailored to advanced LVLMs. Built on the R1-Onevision data foundation, REVERIE+ broadens domain coverage and increases task difficulty, while improving annotation reliability by (i) leveraging multiple models to mine and label diverse negative answers, and (ii) employing a strong commercial LVLM to generate higher-quality reflective rationales. These changes provide richer negative supervision and more accurate reflection signals, making REVERIE+ better suited for hallucination mitigation in more capable models. Extensive experiments across representative hallucination and general multimodal benchmarks show consistent performance gains, validating that the enriched negative supervision and precise reflection signals of REVERIE+ offer a scalable and effective approach for mitigating hallucinations in advanced LVLMs.
