Chang Liu

High-quality light field (LF) acquisition involves a trade-off between spatial and angular resolution. Hybrid camera systems offer a practical solution but present challenges for dense reconstruction due to sparse angular sampling and complex occlusions. To address these issues, this paper proposes a geometry-aware implicit neural representation (INR) framework for LF reconstruction. Distinct from traditional discrete representations or generic ray-space embeddings, we introduce a compact, continuous representation that combines a radiance field with a geometry module, where scalar disparity provides an ideal geometric interpretation and a ray-displacement field serves as the practical realization for occlusion-aware reconstruction. This framework leverages the global epipolar geometry of the 2D LF grid while utilizing the displacement field to correct local geometric inconsistencies caused by occlusions and non-Lambertian effects. By coupling these fields with a radiance network, our method enables end-to-end differentiable optimization from sparse, multi-resolution inputs without relying on large-scale external training datasets. Experiments on hybrid camera data fusion and spatial-angular super-resolution tasks demonstrate that our approach preserves high-frequency details and geometric consistency.

High-fidelity rendering of complex scenes remains challenging for 3D Gaussian Splatting (3DGS), particularly under large viewpoint variations and in regions dominated by high-frequency structures, where aliasing and geometric inconsistencies frequently arise. To address these issues, we propose a wavelet-guided and depth-aware optimization framework that enhances frequency preservation and structural consistency in 3DGS-based computational imaging. A multi-scale wavelet decomposition is integrated to extract discriminative frequency components, which explicitly guide the reconstruction of high-frequency details during splat optimization. In parallel, a depth-aware weighting mechanism is incorporated into the rendering pipeline to adaptively modulate the contributions of spatially adjacent Gaussian primitives, improving geometric stability and mitigating artifacts caused by occlusion or rapid depth transitions. Furthermore, we introduce a high-frequency–weighted hybrid loss function based on L ₁ distance and SSIM, enabling the network to more effectively capture fine structural variations while maintaining perceptual integrity. Experiments on complex indoor and outdoor scenes demonstrate that the proposed method achieves superior detail preservation, reduced aliasing, and improved rendering quality compared with existing 3DGS approaches. These results highlight the potential of frequency- and depth-aware optimization for advancing computational and AI-driven imaging techniques.

Counting small, densely clustered objects from low-altitude aerial views is challenging due to large scale variations, complex backgrounds, and severe occlusion, which often degrade the performance of fully supervised or density-regression methods. To address these issues, we propose a weakly supervised crowd counting framework that leverages point-level supervision and a feature-adaptive fusion strategy to enhance perception under low-altitude aerial views. The network comprises a front-end feature extractor and a back-end fusion module. The front-end adopts the first 13 convolutional layers of VGG16-BN to capture multi-scale semantic features while preserving crucial spatial details. The back-end integrates a Feature-Adaptive Fusion module and a Multi-Scale Feature Aggregation module: the former dynamically adjusts fusion weights across scales to improve robustness to scale variation, and the latter aggregates multi-scale representations to better capture targets in dense, complex scenes. Point-level annotations serve as weak supervision to substantially reduce labeling cost while enabling accurate localization of small individual instances. Experiments on several public datasets, including ShanghaiTech Part A, ShanghaiTech Part B, and UCF_CC_50, demonstrate that our method surpasses existing mainstream approaches, effectively mitigating scale variation, background clutter, and occlusion, and providing an efficient and scalable weakly supervised solution for small-object counting.

Multi-object tracking (MOT) technology integrates multiple fields such as pattern recognition, machine learning, and object detection, demonstrating broad application potential in scenarios like low-altitude logistics delivery, urban security, autonomous driving, and intelligent navigation. However, in open-world scenarios, existing MOT methods often face challenges of imprecise target category identification and insufficient tracking accuracy, especially when dealing with numerous target types affected by occlusion and deformation. To address this, we propose a multi-object tracking strategy based on multi-cue fusion. This strategy combines appearance features and spatial feature information, employing BYTE and weighted Intersection over Union (IoU) modules to handle target association, thereby improving tracking accuracy. Furthermore, to tackle the challenge of large vocabularies in open-world scenarios, we introduce an open-vocabulary prompting strategy. By incorporating diverse sentence structures, emotional elements, and image quality descriptions, the expressiveness of text descriptions is enhanced. Combined with the CLIP model, this strategy significantly improves the recognition capability for novel category targets without requiring model retraining. Experimental results on the public TAO benchmark show that our method yields consistent TETA improvements over existing open-vocabulary trackers, with gains of 10% and 16% on base and novel categories, respectively. The results demonstrate that the proposed framework offers a more robust solution for open-vocabulary multi-object tracking in complex environments.

Inconsistent responses of X-ray detector elements lead to stripe artifacts within the sinogram data, which subsequently manifest as ring artifacts in the reconstructed computed tomography (CT) images, severely degrading image quality. This paper presents a novel method for correcting stripe artifacts in the sinogram data by separating the sinogram into an Ideal Sinogram (IS) and Stripe Artifacts (SA), with both components parameterized through Implicit Neural Representations (INR). The proposed method leverages INR to correct defective pixel response values using implicit continuous functions while simultaneously learning stripe features in the angular direction of the sinogram data. These two components, IS and SA, are combined within an optimization constraint framework, achieving unsupervised iterative correction of stripe artifacts in the projection domain. Experimental results demonstrate that the proposed method significantly outperforms current state-of-the-art techniques in effectively removing ring artifacts while maintaining the clarity and fidelity of CT images, thereby enhancing the overall diagnostic quality of CT imaging.