
Multi-view learning effectively leverages complementary information across heterogeneous feature spaces; however, its performance often deteriorates in extremely low-label scenarios where the scarcity of supervision and varying view quality hinder effective semantic alignment. To address these challenges, we propose CVRF-PLR, a novel multi-view semi-supervised framework that integrates Cross-View Representation Fusion with Pseudo-Label Refinement. Specifically, CVRF-PLR employs an unsupervised reconstruction-based pre-training scheme to derive denoised latent embeddings, effectively filtering view-specific redundancies and noise inherent in raw multi-view data. At its core, a cross-view representation fusion module is designed to capture complex inter-view dependencies through an attention mechanism, further bolstered by pseudo-label-guided semantic enhancement to enforce cross-view semantic consistency. To maximize the utility of abundant unlabeled data under minimal supervision, we develop a $K$-Nearest Neighbor (KNN)-based pseudo-label refinement strategy. By incorporating neighborhood consistency, this strategy generates high-confidence, denoised supervision signals that guide the learning of a discriminative consensus embedding. Extensive experiments on multiple multi-view benchmarks demonstrate that CVRF-PLR significantly outperforms state-of-the-art methods, exhibiting remarkable resilience and superior classification performance in challenging scenarios with extremely sparse labels.
Existing DeepFake detection methods frequently struggle with limited generalization to unseen forgeries, which may be attributed to an over-reliance on dataset-specific semantic shortcuts rather than intrinsic, forgery-specific artifacts. To address this issue, we propose Distribution-Guided Multi- Masked Self-Distillation (DMSD), a framework designed to mitigate the influence of spurious semantic biases through the joint use of distribution-guided stochastic view generation and self-distillation. Specifically, we introduce a Distribution-Guided Multi-complementary Masking (DGMM) strategy, which models the distribution of salient patch representations derived from teacher attention. DGMM resamples salient-token features from this empirical distribution and interpolates them with the original tokens, diversifying dominant prediction evidence while retaining information from each selected token. These distributionresampled patches are then used to dynamically generate multiple complementary masked views, reducing persistent reliance on a fixed subset of local evidence while collectively maintaining tokenposition coverage across views. To effectively leverage the masked views generated by DGMM, we develop a Stochastic Token- Distribution Self-Distillation (STSD) scheme. STSD performs knowledge transfer using features sampled from the empirical teacher token distribution, thereby relaxing deterministic feature matching. By enforcing consistency across multiple stochastically perturbed views, the framework encourages the student to learn invariant and discriminative forgery representations. Extensive experiments demonstrate that DMSD achieves strong and consistent cross-dataset performance across the evaluated benchmarks.
Single-image 3D generative models encode rich geometric priors through large-scale pretraining, making them attractive as foundation models for multi-view reconstruction. However, naively concatenating multi-view DINO features as conditions degrades performance below the single-image baseline, and existing approaches that address this-such as ReconViaGen-fundamentally restructure the conditioning paradigm with VGGT, forcing near-from-scratch retraining and undermining the very priors they aim to exploit. Our key insight is that effectively preserving and exploiting the knowledge learned by the pretrained generative model is important to unlocking its reconstruction capability. To this end, we propose RGB (Reconstruction with Generative Base Model), a framework that introduces cross-view geometric awareness while minimizing disruption to the pretrained generative priors. Specifically, a Multi-View Condition Encoder extracts cross-view geometric information through a dedicated Geometry Encoder decoupled from the generative backbone. Rather than replacing the original semantic representation, it adds the resulting pose-aware view embedding to the DINO semantic tokens, yielding geometry-enhanced condition tokens compatible with the pretrained conditioning interface. We further develop a prior-preserving two-stage training strategy: the Multi-View Condition Encoder is first optimized through the frozen DiT backbone using flow matching and auxiliary geometric supervision; it is then frozen while the DiT backbone is lightly fine-tuned to adapt to the geometry-enhanced conditions without substantially disrupting its pretrained priors. Extensive experiments demonstrate that our method outperforms existing methods, jointly validating the effectiveness and versatility of the proposed framework.
Multimodal large language models (MLLMs) have recently shown promising capability in image quality assessment (IQA) due to their strong semantic understanding. However, most existing MLLM-based IQA methods rely on discrete quality predictions and convert them into mean opinion scores (MOS) through simple aggregation strategies, treating discrete outputs as deterministic scores rather than quantized observations of an underlying continuous quality variable. To address this issue, we propose PGM-IQA, a probabilistic Gaussian mapping framework that transforms calibrated discrete MLLM predictions into continuous perceptual quality distributions. Rather than heuristically aggregating classification scores, PGM-IQA interprets the five ordered quality levels as quantized observations of an underlying continuous quality variable. It first calibrates the class logits through class-specific affine transformations and derives a probability-weighted quality representation from the resulting calibrated categorical distribution over the fixed ordered quality labels. An exact two-output Gaussian process with a coregionalized covariance structure is then introduced to jointly model this intermediate representation and the normalized mean opinion score, thereby capturing both their nonlinear correspondence and statistical dependence. The inferred Gaussian posterior is further projected onto a bounded Beta distribution, yielding a principled predictive distribution over perceptual quality. Its expectation provides the final continuous quality score, while its variance quantifies the associated predictive uncertainty. Extensive experiments on synthetic, real-world, and AI-generated image quality datasets demonstrate competitive and robust performance across diverse evaluation settings.
The number of people in a public place is critical for understanding and analyzing crowd state. If the number of people per frame in an aerial crowd video is directly accumulated, a large number of individuals will be counted repeatedly. Therefore, multi-object tracking (MOT) is a natural solution for inter-frame unduplicated crowd counting. The intrinsic problems in online MOT are tracklet representation and association strategy. However, most existing tracking approaches fail to fully leverage the information across the entire tracklet. To address this issue, we propose a memory-enhanced small-target cluster tracking (MSCT) method, which consists of an Impressed Memory (IMem) network, a memory-based tracklet representation, and a Memory Compensation Matching (MCM) association strategy. Specifically, the IMem network extracts more discriminative individual features using only head-point annotations, without requiring costly bounding box annotations. The memory-based tracklet representation continuously integrates IMem features of targets along a tracklet into several representative memory items, thereby preserving diverse appearance states of the same individual over time. Furthermore, the MCM association strategy leverages memory bias-based compensation instead of relying on explicit motion estimation or global background alignment, effectively addressing unreliable positional predictions. To comprehensively evaluate different approaches, we collect and annotate seven aerial datasets from diverse real-world scenarios. Experimental results show that the proposed method consistently achieves the best performance across all test datasets, with average mAP improvements of 20.40% and 12.90% in hover- and mobile-shooting scenes, respectively, compared with state-of-the-art approaches. The proposed method also demonstrates strong generalization capability across different visual domains in surveillance videos. In addition, the effectiveness of each component is validated through extensive ablation studies. Implementation details will be available at https://github.com/learnsharing/MSCT.