arXiv subjects

Dan Zeng

Publications and source records attributed to Dan Zeng.

At least 19 recordsLinked to original sources

arXiv2026

GeoMAR: Unleashing Geometrically Aligned Features for Masked Autoregressive Blind Face Restoration

Codebook-based blind face restoration (BFR) often suffers from ambiguous conditioning features and a fragile prediction mechanism under severe degradation. To address these challenges, we propose GeoMAR, a framework designed to unleash geometrically aligned features with masked autoregressive (MAR) refinement for robust face restoration. For feature conditioning, we introduce a dual-input extraction pipeline to extract component-based geometric descriptions with explicit, spatially faithful anchors. These textual priors are integrated with low-quality (LQ) features via an Aligned Geometric Priors Injector, which employs a KV-Q exchange strategy to generate geometrically aligned features. For prediction mechanism, we reformulate the one-step mapping into a multi-step MAR process. This coarse-to-fine generation progressively refines complex facial regions based on increasingly reliable context. Experiments on one synthetic and three real-world benchmarks demonstrate that GeoMAR achieves highly competitive perceptual quality and coherent visual structures compared with existing methods. The code is available at https://github.com/BRL-SYSU/GeoMAR.git.

cs.CV↗

arXiv2026

Detail Continuation over a Trustworthy Coarse Scale for Autoregressive Super-Resolution

Hallucination remains a persistent challenge in generative super-resolution (GSR), where reconstructed results may contain visually plausible yet weakly supported content, structural deviations, or unnatural textures with respect to the low-resolution (LR) input. Existing GSR methods have extensively explored the trade-off between perceptual realism and reconstruction fidelity, but the division between preserving reliable coarse-scale information and restoring more uncertain fine details is often handled implicitly within the overall restoration process. Visual autoregressive (VAR) modeling provides a natural opportunity to revisit this issue, as its coarse-to-fine next-scale prediction offers an explicit scale-wise generation interface. However, existing VAR-based SR methods still inherit the original full 1-to-$N$ autoregressive generation path, even though, for super-resolution, coarse-scale information in LR is often relatively more reliable, while long autoregressive chains may accumulate prediction errors. Motivated by these observations, we propose \textbf{K2N}, which reformulates VAR-based SR from full-path generation into a $k$-to-$N$ detail continuation process. Specifically, early coarse-scale states are established directly from LR, while only the remaining finer scales are restored autoregressively. Experimental results show that K2N remains competitive with the VARSR baseline on standard SR metrics, while exhibiting clearer advantages on hallucination-focused evaluation. These findings suggest that explicitly rethinking the generation path in a scale-wise manner can be a promising direction for improving the reliability of generative super-resolution. Our code will be released soon at https://github.com/BRL-SYSU/K2NSR.

cs.CV↗

arXiv2026

VARPose: Flexible 2D Pose Densification via Visual Autoregressive Modeling for Enhanced 3D Lifting

Visual AutoRegressive Modeling (VAR) has excelled in natural image generation via next-scale prediction, but its use on topology-structured data like human skeletons is still unexplored. VARPose is proposed to adaptively densify 2D sparse poses, thereby enriching the anatomical information available for 3D lifting models. Our core contributions are twofold. First, we introduce a Granularity-agnostic Pose Tokenizer (GPT), which employs a single hybrid codebook and a residual quantization strategy to encode poses of varying densities into a unified, multi-scale discrete representation. Our results demonstrate the strong generalizability of this representation. By decoupling the representation from the projection, we can successfully decode novel pose granularities using a frozen codebook with a retrained decoder. Second, we propose UniSkelar, a unified autoregressive model that treats "joint density" as "scale". UniSkelar learns to predict the token sequence for the next density level in a coarse-to-fine manner, conditioned on the sparsest pose. VARPose not only outperforms state-of-the-art methods and generalizes to unseen granularities, but also confers tangible performance gains on downstream tasks, such as 3D Pose Estimation and Human Mesh Recovery, through 2D pose densification. Our code and model are available at https://github.com/BRL-SYSU/VARPose.git.

cs.CV↗

arXiv2026

IPDiff: Diffusion-driven ORSI Salient Object Detection with Information Reconstruction and Multi-Prior Guidance

Existing Salient Object Detection in Optical Remote Sensing Image (ORSI-SOD) methods mainly adopt the static inference strategy, which uses fixed trained model parameters for saliency inference in the testing phase. This means that even if the generated saliency map has errors, it cannot be further optimized. In this paper, we propose the novel IPDiff, a Diffusion-driven ORSI-SOD method with Information Reconstruction and Multi-Prior Guidance. We build IPDiff based on a unique dynamic optimization strategy, which endows IPDiff with the ability to iteratively optimize saliency maps with a dynamic parameter. Specifically, we formulate ORSI-SOD as a conditional diffusion problem in IPDiff. IPDiff first extracts informative conditional priors from ORSIs, including the saliency prior and the hierarchical priors, in the prior network with the assistance of the information reconstruction-driven attention module. The saliency prior can provide positional information of salient objects, while the hierarchical priors can provide specific detail and semantic information of salient objects. Under the guidance of these priors, IPDiff then iteratively denoises random noise as the timestep dynamically changes in the denoising network, generating saliency maps that are close to ground truths. Notably, we simultaneously supervise IPDiff in both spatial and spectral domains through a hybrid loss function to achieve efficient network training. Comprehensive experiments on public ORSSD, EORSSD, and ORSI-4199 datasets demonstrate that our proposed IPDiff achieves the best performance compared to 46 state-of-the-art methods. The code and results of our method are available at https://github.com/MathLee/IPDiff.

cs.CV↗

arXiv2026

Sub-JEPA: Subspace Gaussian Regularization for Stable End-to-End World Models

Joint-Embedding Predictive Architectures (JEPAs) provide a simpleframework for learning world models by predicting future latent representations.However, JEPA training is subject to a bias-variance tradeoff.Without sufficient structural constraints, excessive representationalvariance causes the model to collapse to trivial solutions.The recent LeWorldModel (LeWM) shows that this issue can be alleviated bysimply constraining latent embeddings with an isotropic Gaussian prior.However, latent representations inherently lie on low-dimensional manifoldswithin a high-dimensional ambient space, and enforcing an isotropic Gaussianprior directly in this ambient space introduces an overly strong bias.In this work, we propose ame, which seeks a favorable operatingpoint on the bias-variance frontier by applying Gaussian constraints inmultiple random subspaces rather than in the originalembedding space.This design relaxes the global constraint while preserving itsanti-collapse effect, leading to a better balance between trainingstability and representation flexibility.Extensive experiments across fourcontinuous-control environments demonstrate that consistentlyoutperforms LeWM with very clear margins.Our method is simple yet effective, and serves as a strong baseline for future JEPA-based world model research.fdefinedeeemodeThe code is available at https://github.com/intcomp/Sub-JEPA.

cs.LG↗

arXiv2026

SeGPruner: Semantic-Geometric Visual Token Pruner for 3D Question Answering

Vision-language models (VLMs) have been widely adopted for 3D question answering (3D QA). In typical pipelines, visual tokens extracted from multiple viewpoints are concatenated with language tokens and jointly processed by a large language model (LLM) for inference. However, aggregating multi-view observations inevitably introduces severe token redundancy, leading to an overly large visual token set that significantly hinders inference efficiency under constrained token budgets. Visual token pruning has emerged as a prevalent strategy to address this issue. Nevertheless, most existing pruners are primarily tailored to 2D inputs or rely on indirect geometric cues, which limits their ability to explicitly retain semantically critical objects and maintain sufficient spatial coverage for robust 3D reasoning. In this paper, we propose SeGPruner, a semantic-aware and geometry-guided token reduction framework for efficient 3D QA with multi-view images. Specifically, SeGPruner first preserves semantically salient tokens through an attention-based importance module (Saliency-aware Token Selector), ensuring that object-critical evidence is retained. It then complements these tokens with spatially diverse ones via a geometry-guided selector (Geometry-aware Token Diversifier), which jointly considers semantic relevance and 3D geometric distance. This cooperation between saliency preservation and geometry-guided diversification balances object-level evidence and global scene coverage under aggressive token reduction. Extensive experiments on ScanQA and OpenEQA demonstrate that SeGPruner substantially improves inference efficiency, reducing the visual token budget by 91% and inference latency by 86%, while maintaining competitive performance in 3D reasoning tasks.

cs.CV↗

arXiv2026

STATrack: A Target-Aware Fully Spiking Neural Network for Efficient UAV Tracking

Spiking Neural Networks (SNNs), characterized by their event-driven computation and low power consumption, have shown great potential for energy-efficient visual tracking on unmanned aerial vehicles (UAVs). However, existing SNN-based trackers often rely on costly event cameras, which limits their deployment on standard RGB-camera UAV platforms. To address this limitation, we propose STATrack, a fully spiking neural network framework for UAV visual tracking using only RGB inputs. To the best of our knowledge, this is the first study to explore fully spiking neural networks for RGB-based UAV visual tracking. Considering the challenge of preserving fine-grained target information in deep spiking representations under a limited number of time steps, we introduce Adaptive Mutual Information Maximization (AMIM). AMIM maximizes the statistical dependency between target-centered templates and their deep spiking representations, encouraging the spiking backbone to learn representations that better distinguish the target from surrounding backgrounds. We further develop a sample-difficulty-aware dynamic weighting strategy that adaptively adjusts the mutual-information constraint according to the relative localization difficulty of each training batch. Extensive experiments on four widely used UAV tracking benchmarks demonstrate that STATrack achieves state-of-the-art tracking performance with low theoretical energy consumption, highlighting its potential for energy-constrained UAV applications. Code is released at: https://anonymous.4open.science/r/STATrack.

cs.CV↗

arXiv2026

STENet: Superpixel Token Enhancing Network for RGB-D Salient Object Detection

Transformer-based methods for RGB-D Salient Object Detection (SOD) have gained significant interest, owing to the transformer's exceptional capacity to capture long-range pixel dependencies. Nevertheless, current RGB-D SOD methods face challenges, such as the quadratic complexity of the attention mechanism and the limited local detail extraction. To overcome these limitations, we propose a novel Superpixel Token Enhancing Network (STENet), which introduces superpixels into cross-modal interaction. STENet follows the two-stream encoder-decoder structure. Its cores are two tailored superpixel-driven cross-modal interaction modules, responsible for global and local feature enhancement. Specifically, we update the superpixel generation method by expanding the neighborhood range of each superpixel, allowing for flexible transformation between pixels and superpixels. With the updated superpixel generation method, we first propose the Superpixel Attention Global Enhancing Module to model the global pixel-to-superpixel relationship rather than the traditional global pixel-to-pixel relationship, which can capture region-level information and reduce computational complexity. We also propose the Superpixel Attention Local Refining Module, which leverages pixel similarity within superpixels to filter out a subset of pixels (i.e., local pixels) and then performs feature enhancement on these local pixels, thereby capturing concerned local details. Furthermore, we fuse the globally and locally enhanced features along with the cross-scale features to achieve comprehensive feature representation. Experiments on seven RGB-D SOD datasets reveal that our STENet achieves competitive performance compared to state-of-the-art methods. The code and results of our method are available at https://github.com/Mark9010/STENet.

cs.CV↗

arXiv2026

LightMoE: Reducing Mixture-of-Experts Redundancy through Expert Replacing

Mixture-of-Experts (MoE) based Large Language Models (LLMs) have demonstrated impressive performance and computational efficiency. However, their deployment is often constrained by substantial memory demands, primarily due to the need to load numerous expert modules. While existing expert compression techniques like pruning or merging attempt to mitigate this, they often suffer from irreversible knowledge loss or high training overhead. In this paper, we propose a novel expert compression paradigm termed expert replacing, which replaces redundant experts with parameter-efficient modules and recovers their capabilities with low training costs. We find that even a straightforward baseline of this paradigm yields promising performance. Building on this foundation, we introduce LightMoE, a framework that enhances the paradigm by introducing adaptive expert selection, hierarchical expert construction, and an annealed recovery strategy. Experimental results show that LightMoE matches the performance of LoRA fine-tuning at a 30% compression ratio. Even under a more aggressive 50% compression rate, it outperforms existing methods and achieves average performance improvements of 5.6% across five diverse tasks. These findings demonstrate that LightMoE strikes a superior balance among memory efficiency, training efficiency, and model performance.

cs.LG↗

arXiv2026

PKI: Prior Knowledge-Infused Neural Network for Few-Shot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) aims to continually adapt a model on a limited number of new-class examples, facing two well-known challenges: catastrophic forgetting and overfitting to new classes. Existing methods tend to freeze more parts of network components and finetune others with an extra memory during incremental sessions. These methods emphasize preserving prior knowledge to ensure proficiency in recognizing old classes, thereby mitigating catastrophic forgetting. Meanwhile, constraining fewer parameters can help in overcoming overfitting with the assistance of prior knowledge. Following previous methods, we retain more prior knowledge and propose a prior knowledge-infused neural network (PKI) to facilitate FSCIL. PKI consists of a backbone, an ensemble of projectors, a classifier, and an extra memory. In each incremental session, we build a new projector and add it to the ensemble. Subsequently, we finetune the new projector and the classifier jointly with other frozen network components, ensuring the rich prior knowledge is utilized effectively. By cascading projectors, PKI integrates prior knowledge accumulated from previous sessions and learns new knowledge flexibly, which helps to recognize old classes and efficiently learn new classes. Further, to reduce the resource consumption associated with keeping many projectors, we design two variants of the prior knowledge-infused neural network (PKIV-1 and PKIV-2) to trade off a balance between resource consumption and performance by reducing the number of projectors. Extensive experiments on three popular benchmarks demonstrate that our approach outperforms state-of-the-art methods.

cs.CV↗

arXiv2026

Divide and Conquer: Static-Dynamic Collaboration for Few-Shot Class-Incremental Learning

Few-shot class-incremental learning (FSCIL) aims to continuously recognize novel classes under limited data, which suffers from the key stability-plasticity dilemma: balancing the retention of old knowledge with the acquisition of new knowledge. To address this issue, we divide the task into two different stages and propose a framework termed Static-Dynamic Collaboration (SDC) to achieve a better trade-off between stability and plasticity. Specifically, our method divides the normal pipeline of FSCIL into Static Retaining Stage (SRS) and Dynamic Learning Stage (DLS), which harnesses old static and incremental dynamic class information, respectively. During SRS, we train an initial model with sufficient data in the base session and preserve the key part as static memory to retain fundamental old knowledge. During DLS, we introduce an extra dynamic projector jointly trained with the previous static memory. By employing both stages, our method achieves improved retention of old knowledge while continuously adapting to new classes. Extensive experiments on three public benchmarks and a real-world application dataset demonstrate that our method achieves state-of-the-art performance against other competitors.

cs.CV↗

arXiv2026

Few-shot Class-Incremental Learning via Generative Co-Memory Regularization

Few-shot class-incremental learning (FSCIL) aims to incrementally learn models from a small amount of novel data, which requires strong representation and adaptation ability of models learned under few-example supervision to avoid catastrophic forgetting on old classes and overfitting to novel classes. This work proposes a generative co-memory regularization approach to facilitate FSCIL. In the approach, the base learning leverages generative domain adaptation finetuning to finetune a pretrained generative encoder on a few examples of base classes by jointly incorporating a masked autoencoder (MAE) decoder for feature reconstruction and a fully-connected classifier for feature classification, which enables the model to efficiently capture general and adaptable representations. Using the finetuned encoder and learned classifier, we construct two class-wise memories: representation memory for storing the mean features for each class, and weight memory for storing the classifier weights. After that, the memory-regularized incremental learning is performed to train the classifier dynamically on the examples of few-shot classes in each incremental session by simultaneously optimizing feature classification and co-memory regularization. The memories are updated in a class-incremental manner and they collaboratively regularize the incremental learning. In this way, the learned models improve recognition accuracy, while mitigating catastrophic forgetting over old classes and overfitting to novel classes. Extensive experiments on popular benchmarks clearly demonstrate that our approach outperforms the state-of-the-arts.

cs.CV↗

arXiv2025

SFusion: Energy and Coding Fusion for Ultra-Robust Low-SNR LoRa Networks

LoRa has become a cornerstone for city-wide IoT applications due to its long-range, low-power communication. It achieves extended transmission by spreading symbols over multiple samples, with redundancy controlled by the Spreading Factor (SF), and further error resilience provided by Forward Error Correction (FEC). However, practical limits on SF and the separation between signal-level demodulation and coding-level error correction in conventional LoRa PHY leave it vulnerable under extremely weak signals - common in city-scale deployments. To address this, we present SFusion, a software-based coding framework that jointly leverages signal-level aggregation and coding-level redundancy to enhance LoRa's robustness. When signals fall below the decodable threshold, SFusion encodes a quasi-SF(k +m) symbol using 2^m SFk symbols to boost processing gain through energy accumulation. Once partial decoding becomes feasible with energy aggregation, an opportunistic decoding strategy directly combines IQ signals across symbols to recover errors. Extensive evaluations show that SFusion achieves up to 15dB gain over SF12 and up to 13dB improvement over state-of-the-art solutions.

cs.NI↗

arXiv2025

Motion-Aware Vision-Reference Alignment for Referring Multi-Object Tracking

Referring Multi-Object Tracking (RMOT) extends conventional multi-object tracking (MOT) by introducing natural language references for multi-modal fusion tracking. RMOT benchmarks only describe the object's appearance, relative positions, and initial motion states. This so-called static regulation fails to capture dynamic changes of the object motion, including velocity changes and motion direction shifts. This limitation not only causes a temporal discrepancy between static references and dynamic vision modality but also constrains multi-modal tracking performance. To address this limitation, we propose a novel motion-aware vision-reference alignment framework, named VMRMOT. VMRMOT introduces a motion modality derived from object dynamics to facilitate the alignment between the vision modality and language references. Specifically, we introduce motion-aware descriptions derived from object dynamic behaviors and encode them into motion features as the motion modality through multi-modal large language models (MLLMs). We further design a Vision-Motion-Reference Alignment (VMRA) module to hierarchically align visual queries with motion and reference cues, enhancing their cross-modal consistency. In addition, a Motion-Guided Prediction Head (MGPH) is developed to explore motion modality to enhance the performance of the prediction head. To the best of our knowledge, VMRMOT is the first motion-aware vision-reference alignment framework for the RMOT task. Extensive experiments on multiple RMOT benchmarks demonstrate that VMRMOT outperforms existing state-of-the-art methods. The code is available at https://github.com/Kroery/VMRMOT.

cs.CV↗

arXiv2025

CapeNext: Rethinking and Refining Dynamic Support Information for Category-Agnostic Pose Estimation

Recent research in Category-Agnostic Pose Estimation (CAPE) has adopted fixed textual keypoint description as semantic prior for two-stage pose matching frameworks. While this paradigm enhances robustness and flexibility by disentangling the dependency of support images, our critical analysis reveals two inherent limitations of static joint embedding: (1) polysemy-induced cross-category ambiguity during the matching process(e.g., the concept "leg" exhibiting divergent visual manifestations across humans and furniture), and (2) insufficient discriminability for fine-grained intra-category variations (e.g., posture and fur discrepancies between a sleeping white cat and a standing black cat). To overcome these challenges, we propose a new framework that innovatively integrates hierarchical cross-modal interaction with dual-stream feature refinement, enhancing the joint embedding with both class-level and instance-specific cues from textual description and specific images. Experiments on the MP-100 dataset demonstrate that, regardless of the network backbone, CapeNext consistently outperforms state-of-the-art CAPE methods by a large margin.

cs.CV↗

arXiv2025

Attention Debiasing for Token Pruning in Vision Language Models

Vision-language models (VLMs) typically encode substantially more visual tokens than text tokens, resulting in significant token redundancy. Pruning uninformative visual tokens is therefore crucial for improving computational efficiency, and language-to-vision attention has become a widely used importance criterion for this purpose. However, we find that attention in VLMs is systematically biased. It disproportionately favors tokens appearing later in the sequence, manifesting as over-attention to lower image regions, and assigns inflated scores to semantically empty padding tokens. These behaviors stem from intrinsic recency bias and attention sink effects inherited from large language models (LLMs), and they distort attention-based pruning by preserving irrelevant visual content. To derive a pruning criterion better aligned with semantic relevance, we introduce two lightweight yet effective debiasing techniques that restore the reliability of attention. The first compensates for positional distortions by removing recency-induced attention trends, producing a content-aware and position-agnostic importance measure. The second suppresses attention sink effects by eliminating spurious attention on padding tokens. Our method is model-agnostic, pruning-method-agnostic, and task-agnostic, enabling plug-and-play integration with existing VLM pruning models. Despite its simplicity, our approach consistently delivers strong performance gains. We evaluate our method on ten vision-language benchmarks spanning both image-based and video-based tasks, in comparison with seven state-of-the-art visual token pruning methods and across two representative VLM architectures. Our method achieves substantial performance gains, demonstrating strong effectiveness and generalizability. Our code is available at https://github.com/intcomp/attention-bias.

cs.CV↗

arXiv2025

SMTrack: End-to-End Trained Spiking Neural Networks for Multi-Object Tracking in RGB Videos

Brain-inspired Spiking Neural Networks (SNNs) exhibit significant potential for low-power computation, yet their application in visual tasks remains largely confined to image classification, object detection, and event-based tracking. In contrast, real-world vision systems still widely use conventional RGB video streams, where the potential of directly-trained SNNs for complex temporal tasks such as multi-object tracking (MOT) remains underexplored. To address this challenge, we propose SMTrack-the first directly trained deep SNN framework for end-to-end multi-object tracking on standard RGB videos. SMTrack introduces an adaptive and scale-aware Normalized Wasserstein Distance loss (Asa-NWDLoss) to improve detection and localization performance under varying object scales and densities. Specifically, the method computes the average object size within each training batch and dynamically adjusts the normalization factor, thereby enhancing sensitivity to small objects. For the association stage, we incorporate the TrackTrack identity module to maintain robust and consistent object trajectories. Extensive evaluations on BEE24, MOT17, MOT20, and DanceTrack show that SMTrack achieves performance on par with leading ANN-based MOT methods, advancing robust and accurate SNN-based tracking in complex scenarios.

cs.CV↗

arXiv2025

Open-Vocabulary Camouflaged Object Segmentation with Cascaded Vision Language Models

Open-Vocabulary Camouflaged Object Segmentation (OVCOS) seeks to segment and classify camouflaged objects from arbitrary categories, presenting unique challenges due to visual ambiguity and unseen categories.Recent approaches typically adopt a two-stage paradigm: first segmenting objects, then classifying the segmented regions using Vision Language Models (VLMs).However, these methods (1) suffer from a domain gap caused by the mismatch between VLMs' full-image training and cropped-region inference, and (2) depend on generic segmentation models optimized for well-delineated objects, making them less effective for camouflaged objects.Without explicit guidance, generic segmentation models often overlook subtle boundaries, leading to imprecise segmentation.In this paper,we introduce a novel VLM-guided cascaded framework to address these issues in OVCOS.For segmentation, we leverage the Segment Anything Model (SAM), guided by the VLM.Our framework uses VLM-derived features as explicit prompts to SAM, effectively directing attention to camouflaged regions and significantly improving localization accuracy.For classification, we avoid the domain gap introduced by hard cropping.Instead, we treat the segmentation output as a soft spatial prior via the alpha channel, which retains the full image context while providing precise spatial guidance, leading to more accurate and context-aware classification of camouflaged objects.The same VLM is shared across both segmentation and classification to ensure efficiency and semantic consistency.Extensive experiments on both OVCOS and conventional camouflaged object segmentation benchmarks demonstrate the clear superiority of our method, highlighting the effectiveness of leveraging rich VLM semantics for both segmentation and classification of camouflaged objects.

cs.CV↗