Point cloud denoising is a fundamental yet challenging task in computer graphics. Existing solutions typically rely on supervised training on synthesized noise. However, real-world noise often exhibits greater complexity, causing learning-based methods trained on synthetic noise to struggle when encountering unseen noise-a phenomenon we refer to as noise misalignment. To address this challenge, we propose LaPDA (Latent-space Point cloud Denoising with Adaptivity), a neural network explicitly designed to mitigate noise misalignment and enhance denoising robustness. LaPDA consists of two key stages. First, we adaptively model noise in the latent space, aligning unseen noise distributions with the known training distributions or adjusting them toward distributions with lower noise scales. Training objectives at this stage are formulated based on controlled synthetic noise with varying intensity levels. Second, we introduce a gradual noise removal module that optimizes the spatial distribution of the adaptively adjusted noisy points. Extensive experiments conducted on both synthetic and scanned datasets demonstrate that LaPDA achieves enhanced accuracy and robustness compared to state-of-the-art methods.

Accurate segmentation of cerebral aneurysms from 3D vessel meshes is an essential yet challenging task, complicated by diverse imaging modalities that produce multi-resolution representations. Existing methods often struggle to handle meshes of varying granularity and orientations while maintaining segmentation accuracy. In this paper, we propose an end-to-end multi-scale graph-based segmentation framework that incorporates SE(3)-equivariance and an uncertainty-aware loss function. Our approach constructs a multi-scale graph representation on the 3D mesh and leverages a self-attention mechanism over graph edges to achieve adaptive neighborhood awareness, enabling the network to effectively handle multiple mesh resolutions simultaneously. By introducing an SE(3)-equivariant backbone, rotational variations in aneurysm orientation are naturally accommodated, ensuring relevant and effective feature learning. Furthermore, we develop an uncertainty-aware loss that adaptively emphasizes ambiguous regions, improving segmentation quality and confidence. Experimental results on the public datasets IntrA demonstrate that our method outperforms existing techniques, offering improved accuracy, consistency and stability across different mesh resolutions for 3D aneurysm segmentation. Code is available on https://github.com/Dolphin4mi/se3meshseg.

Unsupervised learning-based action recognition methods have shown immense potential by leveraging vast amounts of unlabeled data, yielding competitive performance in action recognition tasks. To capture semantic information in action, many efforts have focused on multi-stream fusion techniques, which makes the models heavier and less flexible to train. However, existing single-stream methods struggle to provide rich semantic information through data augmentation. To address these challenges, we propose a novel joint-level semantic augmentation method based on a constructed motion manifold, which achieves significant performance gains with virtually no additional training costs. Our approach introduces random perturbations on the motion manifold constructed based on evolutionary metrics. We propose several semantic augmentation strategies, including directional semantic perturbations, magnitude semantic perturbations, and initial pose perturbations. To pay greater attention to samples with significant semantic differences, we introduce the Semantic Adaptive Weighted Loss (SAWL) based on motion manifold distance. SAWL encourages the model to pay more attention to these samples, leading to the learning of an embedding space with semantic invariance. Extensive experiments were conducted on three large-scale datasets, i.e., NTU-60, NTU-120, and PKU-MMD II. The results demonstrate that our single-stream approach, empowered by joint-level semantic augmentation, achieves state-of-the-art (SOTA) performance among single-stream methods and competitive performance among multi-stream methods.

3D human motion prediction is a challenging task due to the highly non-linear nature of movements. Existing deep learning-based action prediction methods emphasize the design of sophisticated network architectures to achieve state-of-the-art performance on datasets. However, in real-life scenarios, changes in skeletal size lead to a shift in data distribution, which presents challenges for accurate motion prediction. Additionally, the presence of stretching artifacts in predicted bone sequences significantly impacts the quality of the predictions. To address these issues, we propose a framework that combines geometric encoding with neural networks to achieve size-invariant and high-quality motion prediction without stretching artifacts. We consider the constraint of bone length and construct a motion space using Riemannian manifold theory, which remains unaffected by changes in skeletal size and can fully represent human motion. Furthermore, we propose the trajectory transport square-root velocity function to encode motion sequences into a flattened space. This transformation simplifies the distance calculation and linearizes the optimization problem in non-flattened space. Experiments on the Human 3.6M and CMU MoCap datasets demonstrated that the proposed method has achieved competitive performance without any stretching artifacts and exhibits robustness to changes in skeletal size.

This paper presents a geometric model for achieving automatic 3D quantitative calculation of the femoroacetabular impingement (FAI) index on computed tomography (CT) images. The result is then used in subsequent work to reduce errors due to perspective differences in existing clinical measurements. This is the first automated quantitative method for a comprehensive FAI diagnostic index that does not rely on datasets. First, the geometric description and equation expression of the hip point cloud were obtained from CT images, fitting the key points and outlines required for the diagnostic index according to such geometric properties as Gaussian curvature. Then, an objective quantitative expression of the FAI index was calculated based on the 3D morphological definition. Next, we statistically analyzed 37 clinical data cases to verify our method’s effectiveness. Especially for the cases of hip dysplasia and acetabular boundary, there were unclear, strong correlations between manual and automatic measures (r = 0.88, ICC = 0.84).