Zunjie Zhu

Simultaneous Localization and Mapping (SLAM) is a crucial technique in computer vision and has been widely used in robot navigation, virtual reality, and augmented reality. Although SLAM algorithms are relatively mature, their performance is severely degraded under low-light conditions. As a core module of SLAM algorithms, feature matching under low-light conditions faces huge challenges. Existing methods train a general feature matching model for all low-light conditions, ignoring the noise modeling for better accuracy. In this work, we introduce LLFeat, a noise-aware method for feature matching under various low-light conditions. To compress the models of various low-light conditions, we further introduce a Noise-aware Feature Modulation (NaFM) layer to the model structure. Therefore, the models can share the most of parameters and preserve only private NaFM layers for each low-light condition, boosting the accuracy of feature matching with negligible additional parameters. LLFeat achieves remarkable results in the highly challenging MID benchmark in both indoor and outdoor scenes, demonstrating the effectiveness of our method. The code will be released.

Video Moment Localization (VML) is a task that seeks to pinpoint the most pertinent segment within an untrimmed video using a linguistic query. Previous works expose the severe data bias issues in VML and note that models avoid understanding visual-textual content by adapting the timestamp distribution. The work investigates data biases from both intrinsic and extrinsic perspectives: The former arises primarily from moment boundary ambiguity and the inputoutput information imbalance. The latter is attributed to the longtail distribution and the semantic bias resulting from the limited tail samples. To reduce the issues, we develop a hybrid multimodal debiasing network with a temporal consistency constraint for VML. Firstly, we propose a multi-temporal Transformer to alleviate boundary ambiguity by merging frame-wise features into segment-wise representations and dynamically aligning with moment boundaries. Subsequently, we implement a temporal consistency constraint to accentuate action information from complex moment context and mitigate the intrinsic bias caused by information imbalance. Moreover, we develop a hybrid linguistic activation module to mitigate the long-tail bias, which offers prior guidance to emphasize distinguishing clues from tail samples. Additionally, we introduce the prior-guided Transformer to alleviate the semantic bias by learning the global semantics of sentences, thereby circumventing the tail-sample overfitting issue. Comprehensive experiments demonstrate the efficacy of our proposed method across three datasets.

This paper introduces a novel and efficient approach for Gaussian Splatting in dynamic scenes that leverages symmetry principles for enhanced computational efficiency and visual fidelity. First, we diverge from conventional methods that process static and dynamic regions uniformly by implementing an adaptive separation mechanism. This approach exploits the inherent symmetry-breaking properties between static and dynamic Gaussian points, utilizing motion differentials to identify and isolate dynamic elements. This symmetry-aware partitioning allows for the application of specialized processing techniques to each region type, with static regions benefiting from their temporal symmetry while dynamic regions receive targeted deformation modeling. Second, through this fine-grained partitioning of static and dynamic components guided by symmetry analysis, we achieve more judicious allocation of computational resources. The symmetric treatment of spatially coherent static regions and the focused processing of symmetry-breaking dynamic elements substantially reduce memory requirements and training time while preserving reconstruction quality. This optimization effectively conserves valuable computational resources without compromising visual fidelity. Third, we introduce a sophisticated deformation modeling framework that learns the transformational characteristics of grids composed of multiple Gaussian points. By incorporating radial basis function principles, which inherently preserve local rotational and translational symmetries, our method efficiently encodes complex motion information of dynamic Gaussian points. This symmetry-preserving deformation approach not only enables high-fidelity reconstruction of dynamic regions but also significantly improves the rendering of continuously evolving shadow interactions by maintaining physical consistency. The result is a marked reduction in visual distortion and rendering outputs that demonstrate exceptional correspondence to ground truth imagery across diverse dynamic scenes.

Neural fields are now the central focus of research in 3D vision and computer graphics. Existing methods mainly focus on various scene representations, such as neural points and 3D Gaussians. However, few works have studied the rendering process to enhance the neural fields. In this work, we propose a plug-in method named K-Buffers that leverages multiple buffers to improve the rendering performance. Our method first renders K buffers from scene representations and constructs K pixel-wise feature maps. Then, We introduce a K-Feature Fusion Network (KFN) to merge the K pixel-wise feature maps. Finally, we adopt a feature decoder to generate the rendering image. We also introduce an acceleration strategy to improve rendering speed and quality. We apply our method to well-known radiance field baselines, including neural point fields and 3D Gaussian Splatting (3DGS). Extensive experiments demonstrate that our method effectively enhances the rendering performance of neural point fields and 3DGS.

The motion capture system that supports full-body virtual representation is of key significance for virtual reality. Compared to vision-based systems, full-body pose estimation from sparse tracking signals is not limited by environmental conditions or recording range. However, previous works either face the challenge of wearing additional sensors on the pelvis and lower-body or rely on external visual sensors to obtain global positions of key joints. To improve the practicality of the technology for virtual reality applications, we estimate full-body poses using only inertial data obtained from three Inertial Measurement Unit (IMU) sensors worn on the head and wrists, thereby reducing the complexity of the hardware system. In this work, we propose a method called Progressive Inertial Poser (ProgIP) for human pose estimation, which combines neural network estimation with a human dynamics model, considers the hierarchical structure of the kinematic chain, and employs a multi-stage progressive network estimation with increased depth to reconstruct full-body motion in real time. The encoder combines Transformer Encoder and bidirectional LSTM (TE-biLSTM) to flexibly capture the temporal dependencies of the inertial sequence, while the decoder based on multi-layer perceptrons (MLPs) transforms high-dimensional features and accurately projects them onto Skinned Multi-Person Linear (SMPL) model parameters. Quantitative and qualitative experimental results on multiple public datasets show that our method outperforms state-of-the-art methods with the same inputs, and is comparable to recent works using six IMU sensors.

Neural fields are now the central focus of research in 3D vision and computer graphics. Existing methods mainly focus on various scene representations, such as neural points and 3D Gaussians. However, few works have studied the rendering process to enhance the neural fields. In this work, we propose a plug-in method named K-Buffers that leverages multiple buffers to improve the rendering performance. Our method first renders K buffers from scene representations and constructs K pixel-wise feature maps. Then, We introduce a K-Feature Fusion Network (KFN) to merge the K pixel-wise feature maps. Finally, we adopt a feature decoder to generate the rendering image. We also introduce an acceleration strategy to improve rendering speed and quality. We apply our method to well-known radiance field baselines, including neural point fields and 3D Gaussian Splatting (3DGS). Extensive experiments demonstrate that our method effectively enhances the rendering performance of neural point fields and 3DGS.

Video compression artifacts arise from quantization applied in the frequency domain. Video quality enhancement aims to reduce such compression artifacts and reconstruct a visually pleasant result. While existing methods effectively reduce artifacts in the spatial domain, they often overlook the rich frequency domain information, especially in addressing multi-scale compression artifacts. This work introduces a frequency-domain upsampling strategy within a multi-scale framework, specifically designed to focus on high-frequency details rather than simply blending neighboring pixels during the upsampling process. Our proposed hierarchical frequency-based upsampling and refinement neural network (HFUR) consists of two modules: implicit frequency upsampling (ImpFreqUp) and hierarchical and iterative refinement (HIR). ImpFreqUp exploits the DCT-domain prior derived through an implicit DCT transform, and accurately reconstructs the DCT-domain signal via a coarse-to-fine transfer. Additionally, HIR is introduced to facilitate cross-collaboration and information compensation between the scales, further refining the feature maps and promoting the visual quality of the final output.We demonstrate the effectiveness of the proposed modules via ablation experiments and visualized results. Experimental results demonstrate that HFUR outperforms the state-of-the-art methods up to 0.13dB/0.17dB on both constant bit rate and constant QP modes. The code is available at https://github.com/zqqqyu/HFUR.

In the domain of cross-view geo-localization, the challenge lies in accurately matching images captured from distinct perspectives, such as aerial drone imagery and satellite imagery of the same geographical location. Existing methods predominantly concentrate on minimizing distances between feature embeddings in the representational space, inadvertently overlooking the significance of reducing embedding redundancy. This oversight potentially hampers the extraction of diverse and distinctive visual patterns critical for precise localization. This work argues that minimizing embedding redundancy is a pivotal factor in enhancing a model’s ability to discriminate diverse scene characteristics. To support this claim, we introduce a straightforward yet effective regularization technique, termed Dynamic Weighted Decorrelation Regularization (DWDR). DWDR serves to actively promote the learning of orthogonal feature channels within neural networks. By dynamically adjusting weights, DWDR targets the minimization of inter-channel correlations, guiding the correlation matrix towards diagonality, indicative of independence among channels. The dynamic weighting mechanism adaptively prioritizes the decorrelation of channels that remain highly correlated throughout training. Additionally, we devise a symmetrical sampling strategy for cross-view scenarios to ensure that the training examples are balanced across different imaging platforms in a batch. Despite its simplicity, the integration of DWDR and the proposed sampling scheme yields remarkable performance across four extensive benchmark datasets: University-1652, CVUSA, CVACT, and VIGOR. Notably, in stringent conditions, such as when constrained to exceedingly compact feature dimensions of 64, our methodology significantly outperforms conventional baselines, thereby affirming its efficacy and robustness under challenging constraints.

In recent years, neural field-based methods for synthesizing novel views have gained popularity due to their exceptional rendering quality and fast training speed. However, the computational cost of volumetric rendering has significantly increased with the advancement of camera technology and the subsequent rise in average camera resolution. Despite extensive efforts to accelerate the training process, the training duration remains unacceptable for high-resolution inputs. Therefore, it's crucial to develop efficient sampling methods to optimize the learning process of neural fields from large inputs. In this paper, we present a new technique called Superpixel-based Efficient Sampling (SES) to improve the learning efficiency of neural fields. Our approach optimizes pixel-level ray sampling by segmenting the error map into multiple superpixels and dynamically updating their errors during training to increase ray sampling in superpixel areas with higher rendering errors. Compared with other methods, our approach leverages the flexibility of superpixels, effectively reducing redundant sampling while considering local information. Our method not only speeds up the learning process but also enhances the rendering quality learned from large inputs. We conduct extensive experiments to evaluate the effectiveness of our method across several baselines and datasets. The code will be released.