ABSTRACT Real‐time photorealistic novel view synthesis of dynamic scenes remains a challenging task, primarily due to the inherent complexity of temporal dynamics and motion patterns. Despite recent methods based on Gaussian Splatting having shown considerable progress in this regard, they are still limited by high memory consumption. In this paper, we propose a time‐varying 3D Gaussian splatting (TVGS) representation for dynamic scenes, which incorporates two key components. First, we model the scene using 3D Gaussians endowed with temporal opacity and time‐varying motion parameters. These attributes effectively capture transient phenomena such as the sudden appearance or disappearance of dynamic elements. Second, we introduce an adaptive density control mechanism to optimise the distribution of these time‐varying 3D Gaussians throughout the sequence. As an explicit dynamic scene representation, TVGS not only achieves high‐fidelity view synthesis but also attains a real‐time rendering speed of 160 FPS on the Neural 3D Video Dataset using a single RTX 4090 GPU.

ABSTRACT Panoramic image stitching algorithms inevitably introduce distortions due to their inherent limitations. Due to the characteristics of the Human Visual System (HVS), slight stitching distortions have a minimal impact on perceived image quality, whereas severe distortions significantly degrade it. To address this issue, a visual perception‐based quality assessment method for stitched panoramic images is proposed. Existing no‐reference image quality assessment (NR‐IQA) methods often struggle to accurately model the HVS and its subjective perception of stitching distortions, leading to significant discrepancies between predicted and actual quality scores. Therefore, we use the learnt perceptual image patch similarity (LPIPS) to capture perceptual similarity from the HVS, thereby generating a high‐quality pseudo‐reference image (PRI). To further improve its perceptual quality, the PRI is refined using a visual restoration network (VRN) to reduce residual distortions. Subsequently, both the pseudo‐reference and distorted images are fed into a quality assessment network with shared parameters to extract deep feature representations. To improve distortion‐aware quality evaluation, we design an Adaptive Feature Fusion (AFF) module, where features from the pseudo‐reference and distorted images are first enhanced through an attention mechanism and then adaptively fused via learnt weights for more effective quality assessment. Finally, the fused features are processed through fully connected layers for regression, predicting the perceptual quality score of the stitched panoramic image. Extensive experimental results demonstrate that the proposed method outperforms state‐of‐the‐art approaches, achieving significant improvements across multiple evaluation metrics.

This paper presents a new method for calculating similarity. It's called the unit dot product similarity method. The proposed method, unlike the traditional dot product method, can maintain the similarity of equally scaled vectors and get a bounded similarity results. We develop and compare the proposed method in the attention-based encoder-decoder structure. The proposed method brings further improvement to the recognition results. For the end-to-end speech recognition model, we select greedy soup instead of the average model parameters in WeNet. We proposed a dynamic parallel greedy soup optimization algorithm to increase computational speed. The experiments show the importance of proposed method and optimization algorithm. The effectiveness is also proved on multiple corpora.

Deep learning excels in natural image object detection, but remote sensing images face challenges like multidirectional objects and neighborhood interference. Existing methods use shared features for classification and regression, causing task interference. Classification needs translation/rotation-invariant features, while regression requires translation/rotation-equivariant features. Additionally, regression parameters (e.g., center, shape, and angle) demand distinct feature properties. To address this, we propose TSFD-Net, featuring: 1) task differential decoupling module (TDDM): decouples task-specific features via parallel CNN-Transformer branches, and 2) parameter differential decoupling module (PDDM): designs specialized regressors for distinct parameters (e.g., angle versus center/shape). Together, TDDM and PDDM form the two-stage feature decoupling (TSFD) structure. We further introduce dynamic cascade activation masks (DCAMs), leveraging bounding box feedback to enhance target focus and suppress neighborhood noise. TSFD network (TSFD-Net) achieves state-of-the-art results on DOTA-v1.0 (81.37% mAP), validating its efficacy.

Neural radiance fields (NeRF) and its variants have achieved remarkable success for novel view synthesis. Most existing radiance field models utilize the mean squared error (MSE) as the photometric loss, which is prone to resulting in blurriness and geometry inaccuracy, especially for sparse views. Instead of the pixel-wise loss, we introduce the Pearson correlation coefficient (PCC) for constructing a new photometric loss from the perspective of linear correlation. Due to the relativeness of PCC, we rectify PCC to absolutize it. To be specific, we relax the denominator of PCC based on the inequality of arithmetic and geometric means to enforce unit scale, and add an extra modulation factor to further enforce zero location. The experimental results show the proposed loss is significantly better than the MSE loss, e.g. the peak signal-to-noise ratio (PSNR) increasing by 186% for TensoRF on Replica dataset, and 3∼5dB for DVGO at scenes from Tanks and Temples dataset with sparse views setting.

To reduce key disagreement rate and increase key generation rate, this paper proposes a lightweight and robust shared secret key extraction scheme from atmospheric optical wireless channel. A conception of grouping samples by using B+ tree and quantizing groups is introduced to improve the balance of distribution of channel measurement values in quantization intervals. Based on the conception, a basic B+ tree quantization algorithm (B-BPTQ) and a compensation-technology-based B+ tree quantization algorithm (C-BPTQ) are designed. A LDPC-based multi-scale information reconciliation algorithm (LDPC-MSIR) is put forward to simplify information reconciliation and reduce its computation complexity. An enhanced randomness post-processing algorithm is formulated to increase statistical randomness of initial key sequence. Finally, the performance of the proposed scheme is validated by numerical simulations and field experiments. The experimental results show that it has better performance in terms of key disagreement rate, key generation rate and key randomness compared with other schemes.

Abstract The quality of images generated by computer graphics rendering algorithms is mainly affected by visible distortion at some pixel locations. Image quality assessment (IQA) metrics are commonly utilized to assess the quality of rendered images, but their results are a global difference value, which does not provide pixel‐wise differences to optimize the renderings. In contrast, visibility difference models including visual perception models and deep learning models can calculate pixel‐wise visibility difference between distorted images and reference images. However, they either are only applied to a single type of visible distortion or are seriously dependent on datasets. To this end, the authors propose a novel model, dubbed Human Visual Perception and Deep Learning Image Difference Metric (HPDL‐IDM), which combines the Human Visual System (HVS) model and deep learning. HPDL‐IDM primarily consists of two modules: (i) the visual perception feature calculation module, which calculates difference maps between various kinds of features extracted from the reference image and the distorted image according to the visual characteristics of human eyes and concatenates them, and (ii) the deep learning module, which utilizes a neural network of encoder–decoder structure to train on the LocvisVC and VisTexRes datasets whose input and output are these concatenated feature difference maps and the final image distortion visibility difference map respectively. Additionally, the authors pool the final difference map into a global difference value between 0 and 1 to apply their model to many image processing tasks related to Image quality metrics (IQMs). Experimental results show that HPDL‐IDM's generalization capacity and accuracy are improved by a large margin compared to other models.

Real-time 3D reconstruction of static scenes can be achieved based on the RGB-D image sequence fusion. It is a popular practice to divide the space into uniform voxels and use a truncated signed distance function to represent surface information. In order to represent a scene of large scale, the voxel hash algorithm which stores voxels compressively can be used, but most of the conventional methods do not consider the complexity and roughness of the object surface in the scene, so the scene is represented with a uniform resolution. It somewhat limits the range of scene representation and the speed of real-time reconstruction. In this paper, a large-scale scene reconstruction algorithm based on voxel hashing storage with LOD representation is proposed. The main contributions include the following two aspects: (1) By preprocessing the depth image with smooth filtering, which ensures the accuracy of the data, it can effectively reduce the distortion caused by the sensor itself and violent motion and provide better support for the stages of voxel hashing, model rendering, and frame-to-model camera position tracking. (2) The 3D reconstruction with LOD representation is realized. We take the view distance and the roughness of the model surface as criteria to control the adaptive division and representation of spatial voxel blocks. Finally, we carried out qualitative and quantitative evaluations of the algorithm, and confirmed that the algorithm can achieve real-time reconstruction with different levels of detail in the commercial graphics processing hardware environment, and achieve a good fusion effect in large-scale scenes.

Vector quantization is a viable solution to the problem as to full utilization of randomness provided by statistically dependent channel measurements in secret key extraction from randomly fluctuating channels. By aiming at atmospheric optical wireless channels, new vector quantization schemes based on sample grouping (SG) are proposed by using different ways, specifically, the$K\text {-dimensional}$tree (KDT) partition and random placement, to separate samples of the dimension-reduced channel coefficient vector into a given number of groups, with each containing equal number of samples. Use of the KDT partition leads to both the basic KDT-partition-based quantization (B-KDTPQ) and improved KDT-partition-based quantization (I-KDTPQ) schemes. On the other hand, utilization of the random placement brings forth the random equisized grouping based quantization (REGQ) scheme. Performance evaluation shows that, due to use of quantization symbol modification, the I-KDTPQ scheme always has lower quantization symbol disagreement rate (QSDR) than the B-KDTPQ scheme; compared with the I-KDTPQ scheme, the REGQ scheme has more flexibility in accommodating itself to various conditions of channel coefficient fluctuations and noise levels.