Light field angular super-resolution (LFASR) aims to synthesize dense novel views from sparsely sampled light fields, thereby enabling immersive applications such as virtual and augmented reality. However, existing methods primarily focus on 2D cross-view interactions while insufficiently leveraging 3D scene geometry, leading to reduced angular consistency and suboptimal view synthesis-particularly in complex scenes. To address this limitation, we present an implicit-explicit hybrid representation and implement it within a geometry-aware implicit-explicit framework (GIENet) that enhances geometric reasoning for LFASR. Specifically, we propose an implicit geometry-aware interaction module, inspired by multi-plane image (MPI) representations, which captures geometry-aware features through structured interactions in a layered representation space. To overcome the representational constraints of a finite MPI stack, we further introduce an explicit geometric feature enhancement module that integrates self-inferred geometric priors into the reconstruction pathway. In addition, we propose an axis-aligned multi-plane aggregation module that fuses multi-view features along axis-aligned planes using separable convolutions, achieving both computational and memory efficiency while preserving fine details. Extensive experiments demonstrate that GIENet not only achieves superior reconstruction quality but also offers competitive computational efficiency and remarkable robustness across diverse upsampling factors and datasets.

Light field imaging, with its ability to record both the intensity and direction of light rays, has enabled realistic simulation of focus distance variations for refocusing and novel view synthesis. However, traditional light field imaging technology depends on dense sampling, making it difficult to attain high spatial and angular resolutions simultaneously. To address these limitations, we propose hybrid light field (HyLF), a self-supervised framework that realizes sparse angular super-resolution without requiring prior knowledge of camera parameters. HyLF integrates an implicit neural light field with an explicit voxel representation to achieve high-quality angular super-resolution from sparse inputs. First, we propose a jointly optimized two-plane embedding strategy that effectively encodes each ray in the absence of camera parameters. Next, an implicit neural multiplanar light field models focus distance variations with embedded rays, extracting more features to address the challenge of sparse angular sampling. Finally, an explicit planar-prior voxel representation decomposes the multiplane features predicted by the implicit module into voxel grids across focus planes, isolating salient voxels to enhance angular super-resolution. Comprehensive experiments demonstrate that HyLF outperforms existing baselines, achieving higher quantitative metrics, clearer visual quality, and faster inference speed.

Light field (LF) cameras record both intensity and directions of light rays in a scene with a single exposure. Due to the trade-off between spatial and angular dimensions, the spatial resolution of LF images is limited, so super-resolution is widely studied. Pixels follow linear coordinate projection across views in LF images. Hence, auxiliary views nearer to the target view are generally more effective for use in super-resolution. In this paper, an LF-pyramid is proposed based on an angular-distance constraint for discriminatively exploiting auxiliary views. From views of different layers in an LF-pyramid, complementary features of different effectiveness can be extracted. However, shapes of LF-pyramids change for target views with different angular positions. To fully exploit an LF-pyramid, we introduce a pyramid-angular-constraint network for LF super-resolution (LF-PACNet). Specifically, to handle an arbitrary number of views in each layer, an intra-pyramid-layer feature extraction module is designed, which treats all views in the same layer equally in complementary information extraction. Then, to deal with an arbitrary number of layers, a recurrent cross-pyramid-layer feature complementation module is constructed, which discriminatively complements the target view with high-frequency details. Extensive experiments on public datasets demonstrate state-of-the-art performance for our method, both visually and numerically, especially for datasets with large disparities.

Various deep learning-based light field image super-resolution methods have attained notable success in recent years. However, most of them focus on encoder design while neglecting the critical role of upsampling process in decoder part. Motivated by the recent progress in single image domain with implicit neural representation, we elaborately propose a spatial-angular-epipolar implicit image function (SAEIIF) in this paper, which can redefine the upsampling process to significantly improve performance and enable arbitrary-scale light field super-resolution. Specifically, it contains two complementary upsampling branches. One branch incorporates spatial implicit image function (SIIF) and angular implicit image function (AIIF) to mine intra-view information in sub-aperture images and inter-view information in macro pixels. The other branch involves epipolar implicit image function (EIIF) to leverage spatial-angular correlation in epipolar plane images. By decomposing SIIF, AIIF and EIIF into horizontal and vertical two-step upsampling to form a perfect match of upsampling scale, SAEIIF introduces a multi-stage feature interaction architecture across two branches to fully merge spatial, angular and epipolar domain information. Furthermore, we optimize feature sampling strategy based on characteristics of sub-aperture images, macro pixels, and epipolar plane images, introducing horizontal-vertical separable local sampling for SIIF and AIIF, as well as dual-source oriented line sampling used for EIIF. The extensive experimental results demonstrate that our SAEIIF can be effectively integrated with most encoders and achieve outstanding performance on both fixed-scale and arbitrary-scale light field spatial super-resolution, angular super-resolution, spatial-angular joint super-resolution.

Current research on radar-camera fusion is predominantly oriented toward vehicle-mounted applications, thereby leaving a critical research gap in the development of adaptive fusion solutions for complex roadside perception scenarios. Comprehensive analysis indicates that existing vehicle-centric approaches fail to fully account for the inherent heterogeneities of fixed roadside infrastructures, including the fixed geometric configuration of sensors and the nonlinear degradation of sensor reliability across large-scale perception domains. To address this research gap, this study proposes a novel Spatio-Temporal Road Structure-aware Radar-camera Fusion (ST-RSRF) framework. First, the ST-RSRF framework leverages invariant road structure information as a deterministic geometric constraint for target data association, effectively mitigating radar under-segmentation and cross-lane matching ambiguities in dense traffic environments. Second, an adaptive trajectory fusion strategy is developed by integrating a spatio-temporal weighting model into the classical Kalman filter framework. This model dynamically adjusts the weighting of sensor observations based on target distance and time-varying environmental factors (e.g., illumination intensity and visibility conditions) to sustain optimal state estimation performance. Real-vehicle experiments were conducted at a dedicated test facility to validate the superiority of the proposed ST-RSRF framework. In terms of target matching performance, ST-RSRF achieves a 100% Matching Rate (MR) in small-target parallel driving and stationary interference scenarios, whereas traditional Kuhn–Munkres (KM) algorithms only attain 60% and 25%, respectively. For trajectory fusion, quantitative evaluation against ground-truth detection lines demonstrates that ST-RSRF reduces the Mean Absolute Error (MAE) by 7.85%, the Root Mean Square Error (RMSE) by 4.86%, and the Median Error by 24.80%. Collectively, these experimental results illustrate that ST-RSRF significantly enhances the accuracy and environmental robustness of roadside perception systems under diverse all-weather operating conditions.

The Light Field (LF) captures both spatial and angular information of scenes, enabling precise depth estimation. Recent advancements in deep learning have led to significant success in this field; however, existing methods primarily focus on modeling surface characteristics (e.g., depth maps) while overlooking the depth space, which contains additional valuable information. The depth space consists of numerous space points and provides substantially more geometric data than a single depth map. In this paper, we conceptualize depth prediction as a spatial modeling problem, aiming to learn the entire depth space rather than merely a single depth map. Specifically, we define space points as signed distances relative to the scene surface and propose a novel distance-constraint query mechanism for LF depth estimation. To model the depth space effectively, we first develop a mixed sampling strategy to approximate its data representation. Subsequently, we introduce an encoder-decoder network architecture to query the distances of each point, thereby implicitly embedding the depth space. Finally, to extract the target depth map from this space, we present a generation algorithm that iteratively invokes the decoder network. Through extensive experiments, our approach achieves the highest performance on LF depth estimation benchmarks, and also demonstrates superior performance on various synthetic and real-world scenes.

Visual tracking that combines RGB and thermal infrared modalities (RGB-T) aims to utilize the useful information of each modality to achieve more robust object localization. Most existing tracking methods based on convolutional neural networks (CNNs) and Transformers emphasize integrating multi-modal features through cross-modal attention, but ignore the potential exploitability of complementary information learned by cross-modal attention for enhancing modal features. In this paper, we propose a novel hierarchical progressive fusion network based on cross-modal attention guided enhancement for RGB-T tracking. Specifically, the complementary information generated by cross-modal attention implicitly reflects the consistent regions of interest of important information between different modalities, which is used to enhance modal features in a targeted manner. In addition, a modal feature refinement module and a fusion module are designed based on dynamic routing to perform noise suppression and adaptive integration on the enhanced multi-modal features. Extensive experiments on GTOT, RGBT234, LasHeR and VTUAV show that our method has competitive performance compared with recent state-of-the-art methods.

Light field imaging provides rich geometric information by simultaneously capturing spatial and angular dimensions, enabling precise depth estimation which is crucial for numerous vision tasks. Existing volume-based depth estimation methods primarily focus on the center view, ignoring consistency across full views due to insufficient angular coherence and computational complexity. To address these limitations, we propose a novel difference-guided full-view volume framework, which integrates a full-view difference volume with a center-view matching volume. Specifically, the full-view difference volume exploits angular diversity, providing robust angular correspondences that help address ambiguities in low-texture regions and improve consistency across views. The center-view volume preserves precise spatial matching correspondences that are essential for accurate estimation. By propagating the center density information across all views and combining them in final rendering, proposed method effectively leverages two complementary correspondences. Furthermore, we introduce an assisted edge-aware module that leverages multi-view pixel visibility to modulate feature weights. It mitigates occlusions and enforcing edge consistency across all views. Finally, we propose a difference-guided full-view density propagation network for light field full-view depth estimation. Experimental results on synthetic and real light field datasets demonstrate the effectiveness of our method in simultaneously estimating high-quality depth maps for all views. Compared to the previous full-view and arbitrary-view models, our framework achieves over a 8% reduction in Mean Square Error and more than a 3% reduction in BadPixel (7). Furthermore, the proposed approach realizes a 3 ×  to 40 ×  speedup in inference time relative to existing techniques.

3D Gaussian Splatting (3DGS) has recently demonstrated remarkable performance in novel view synthesis. Nevertheless, existing 3DGS approaches typically rely on dense multi-view supervision and known camera pose priors, leading to substantial performance degradation under sparse-view settings or when pose information is unavailable. To overcome these limitations, we propose a pose-free sparse-view 3DGS framework enhanced by semantic-regularized optimization. Specifically, to relax the reliance on known poses, we jointly optimize camera poses-initialized by a feed-forward network-together with the Gaussian splats. Moreover, to alleviate the challenges posed by sparse-view synthesis, we incorporate semantic regularization, which complements depth cues and provides an additional constraint on the 3DGS representation. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art novel view synthesis baselines, while simultaneously reducing geometric distortions and enhancing visual fidelity.

Federated learning enables collaborative model training among multiple parties without exposing raw local data. Decision trees are especially suitable for tabular data due to their interpretability and efficiency, yet existing privacy-preserving horizontal federated decision tree methods often suffer from heavy communication overhead and expensive cryptographic operations. In this paper, we propose a communication-efficient two-party secure decision tree framework for multi-class classification in horizontal federated learning. By reformulating the Gini impurity evaluation into a matrix-based secure computation architecture, our method parallelizes complex non-linear operations and greatly reduces repeated scalar interactions. We further design a suite of lightweight secure matrix protocols based on additive secret sharing and construct a secure recursive tree-building pipeline under the semi-honest adversary model. Experimental results on five public datasets demonstrate that the proposed method achieves competitive classification accuracy while exhibiting favorable efficiency, especially for datasets with large sample sizes.

Light field (LF) full-view depth estimation aims to recover dense and coherent depth maps for all sub-aperture views, which is crucial for applications such as 3D reconstruction, LF editing and virtual reality. However, directly extending center-view volume-based methods to the full-view is computationally infeasible, as it requires constructing a separate cost volume for each view. Besides, existing full-view propagation-based approaches, while more efficient, frequently suffer from edge fattening and cross-view inconsistencies in the presence of occlusions. In this paper, we propose a gradient-guided density redistribution network (GDRNet), a novel end-to-end framework that efficiently generates full-view depth maps by constructing a single plane-density volume and a multi-plane depth image, which are then propagated to all angular views. To resolve ambiguous estimates at occlusion edges, we perform a direction-aware gradient-guided density redistribution only inside a dilated edge narrow band. For each center pixel in edge regions, a guidance gradient is derived from the initial depth map to determine the normal and tangent directions. Then, density in edge fattening regions can be redistributed via sampling along the normal direction, while similarity along the tangent direction can fill bad pixels with inconsistencies. Furthermore, an adaptive edge extraction module with four directional learnable Sobel kernels is designed to jointly exploit spatial and angular gradients, enabling robust detection and localizing the refinement band. Extensive experiments on synthetic and real-world LF datasets demonstrate that GDRNet achieves state-of-the-art accuracy and edge sharpness in both quantitative and qualitative evaluations, while maintaining computational efficiency compared to full-view methods.

Achieving unified scene parsing that simultaneously outputs cross-domain semantic segmentation and depth estimation without scene-specific retraining is crucial for robust perception in complex real-world environments, yet remains a challenging goal. While recent monocular depth estimation models such as DepthAnything V2 exhibit strong domain generalization, semantic segmentation still suffers from severe structural degradation under domain and viewpoint shifts. We observed a persistent hierarchical calibration gap, where Euclidean representations exhibit larger calibration gaps between child and parent categories under domain shifts, suggesting limitations of existing Euclidean-based methods in preserving semantic hierarchies. To address this issue, we propose HyperMapper, a hyperbolic structure-aware mapping framework that bridges semantic understanding and geometric priors through hyperbolic token-to-feature interactions. By exploiting the negative curvature of hyperbolic space, HyperMapper helps capture hierarchical relationships and maintains geometric consistency across domains. Furthermore, by combining the expressive priors of vision foundation models (VFMs) with parameter-efficient fine-tuning (PEFT), HyperMapper achieves cross-domain adaptation with minimal trainable parameters in backbone while retaining the strong depth estimation capability of DepthAnythingV2 without retraining. Extensive experiments on multiple cross-domain and cross-viewpoint benchmarks demonstrate that HyperMapper achieves a higher mIoU for both parent and child categories while consistently improving segmentation accuracy over strong baselines. Our approach establishes a promising direction for task-preserving dual-task adaptation, bridging semantic and geometric learning and paving the way toward unified, cross-domain scene parsing.

With the development of Connected Vehicles technology and Cooperative Vehicle Infrastructure System, the "long" platoon has become a promising trend of platooning technology, and long platoons can take the full advantages of platoons in enhancing traffic efficiency and reducing energy consumption. In this paper, we propose a Cellular Automata-based long platoon model in which the platoon is divided into several sub-platoons and virtual leading vehicles are assigned to the sub-platoons dynamically according to the surrounding traffic states. Moreover, to evaluate the proposed model, it is compared with the Lenarska's model and the traditional Cooperative Adaptive Cruise Control (CACC) model by simulations, and the influences of the long platoon size and traffic perturbations on the platoon are analyzed. The simulations indicate that for the acceleration and deceleration perturbation scenarios, the virtual leaders effectively divide the long platoon into multiple sub-platoons, and its sequence can change dynamically to reduce the influence of the perturbation on the platoon. Compared to the Lenarska's model and the CACC model, the proposed model reacts to the speed perturbations faster and has smaller speed variations. The proposed model has better stability and safety and is more efficient than the Lenarska's model and the CACC model.

Multi-object tracking (MOT) has witnessed significant advancements in recent years, yet it remains challenged by complex uncertainties arising from pedestrian movement patterns. To address this, we present a unified framework that explicitly models pedestrian dynamics through a dual-phase paradigm, combining a Hidden Markov Model (HMM) for motion modeling and weighted entropy for adaptive multi-cue fusion. Furthermore, a multi-agent architecture is employed for track management, enabling parallelized state estimation and seamless integration of the HMM-based Kalman filter with multi-cue fusion. Quantitative evaluations show that our method achieves 82.1 in IDF1, 81.5 in MOTA, 65.9 in HOTA, and 1,255 IDs on the MOT17 benchmark, and achieves 81.2 in IDF1, 78.4 in MOTA, 65.7 in HOTA, and 608 IDs on the MOT20 benchmark, and the application of the multi-agent mechanism significantly improves the scores on FPS as a result of efficient computation. The experimental results demonstrate that the proposed method achieves state-of-the-art performance, particularly in highly crowded scenes.

Light field (LF) encodes both intensity information and directional information of all light rays into high-dimensional signal, which facilitates various advanced applications due to its rich description. However, current mainstream research adopts two-plane parametrization to describe 4D LF, losing the information stored in the spectral dimension that can delineate more scene details. On this account, we introduce 5D hyperspectral light field (H-LF) to achieve robust semantic segmentation for the first time. To alleviate data redundancy while preserving useful information to a large extent, we use pseudo H-LF with sparsely non-repetitive angular-spectral distribution as an alternative and propose a network called PHLFNet. Specifically, our network successively performs feature-level angular-spectral joint blending and semantic-level angular-spectral joint enhancement to fully exploit the complementary information embedded in pseudo H-LF, in which the former executes preliminary information fusion and calibration across all modalities, and the latter distills unique semantic cues of each auxiliary modality to boost feature of segmented central view image. To guarantee the accuracy of semantic cues distillation, we design boundary consistency semantic label propagation to handle cross-spectral color inconsistency and cross-angular pixel misalignment in pseudo H-LF, thereby generating semantic labels of each auxiliary modality to provide supervision. Extensive experimental results illustrate that PHLFNet achieves outstanding performance compared with relevant state-of-the-art methods, demonstrating the significance of introducing H-LF for semantic segmentation.

Compared to supervised learning methods, self-supervised learning methods address the domain gap problem between light field (LF) datasets collected under varying acquisition conditions, which typically leads to decreased performance when differences exist in the distribution between the training and test sets. However, current self-supervised light field angular super-resolution (LFASR) techniques primarily focus on exploiting discrete spatial-angular features while neglecting continuous LF information. In contrast to previous work, we propose a self-supervised unconstrained neural light field (UNeLF) to continuously represent LF for LFASR. Specifically, any LF can be described as the camera pose for each sub-aperture image (SAI) and the two-plane that captures these SAIs. To describe the former, we introduce a SAIs-dependent pose optimization method to solve the issue that arises from the narrow baseline of most LF data, which hinders robust camera pose estimation. This mechanism reduces the number of trainable camera parameters from a quadratic to a constant scale, thereby alleviating the complexity of joint optimization. For the latter, we propose a novel adaptive two-plane parameterization strategy to determine the two-plane that captures these SAIs, facilitating refocusing. Finally, we jointly optimize the camera parameters, near-far planes and neural light field, efficiently mapping each adaptive two-plane parameterized ray to its correspondence color in a continuous manner. Comprehensive experiments demonstrate that UNeLF achieves faster training and inference with fewer computational resources while exhibiting superior performance on both synthetic and real-world datasets.

This report summarizes the 3rd NTIRE challenge on light field (LF) image super-resolution (SR), focusing on novel methods and their outcomes. This challenge aims to super-resolve LF images degraded by bicubic downsampling, and comprises three tracks: a classical track, an efficiency track, and a large model track. In total, 308 participants registered, and 13 teams submitted results that outperformed the baseline methods. The challenge has established a new state-of-the-art in LF image SR, e.g., the winning method in Track 1 achieves a 0.36 dB PSNR improvement over last year's champion on the test set. We present the submitted solutions, analyze their common trends, and highlight practical techniques. We hope this challenge will inspire further advancements in LF image SR.

Light field (LF) technology captures both spatial and angular information of the real world, enabling accurate depth estimation. Cost volume-based methods mostly consider LF depth estimation as a shift-matching process, which fail to efficiently establish the relationship among different viewpoints. State Space Model (SSM) has shown strong capabilities in long-sequence modeling, providing a powerful mechanism to capture viewpoints associations. In this paper, we observe that LF depth estimation can be viewed as state transition and then propose a text-similar representation based on the distribution of pixel values across different viewpoints, which is able to detect occluded and discontinuous regions. Furthermore, to extract the potential depth features, we represent it as Depth State Space Model (DSSM), leveraging the state transition mechanism of SSM to capture spatial, angular and structural characteristics in complex regions. Based on the proposed DSSM, we develop DSS-Net for depth estimation. Experiments demonstrate that our approach achieves state-of-the-art performance, with significant improvements in occluded and discontinuous regions, highlighting its effectiveness in addressing the complexities of LF depth estimation.

Dense sampling of the light field (LF) is essential for various applications, such as virtual reality. However, the collection process is prohibitively expensive due to technological limitations in imaging. Synthesizing novel views from sparse LF data, known as LF Angular Super-Resolution (LFASR), offers an effective solution to this problem. Accurate cross-view interaction is crucial for this task, given the complementary information between LF views. Previous methods, however, suffer from limited reconstruction quality due to inefficient view interaction. To address this, we propose a Multiplane-based Cross-view Interaction Mechanism (MCIM) for robust LFASR. Extensive comparisons with state-of-the-art methods demonstrate that our method achieves superior performance, both visually and quantitatively. Specifically, Drawing inspiration from MultiPlane Images (MPI) in scene modeling, our mechanism incorporates a novel Multiplane Feature Fusion (MPFF) strategy. This strategy facilitates fast and accurate cross-view interaction, enhancing the network’s robustness to scene geometry and suitability for different-baseline LF scenes. Furthermore, to address information redundancy in multiplanes, we leverage the transparency property of MPI and devise a plane selection strategy. Finally, we propose CSTNet, a Cross-Shaped Transformer-based network for LFASR, which employs a cross-shaped self-attention mechanism to enable low-cost training and inference. Experimental results on various angular super-resolution tasks validate that our network achieves state-of-the-art performance on both synthetic and real-world LF scenes.