Tag
torch-pointcloud 是一个遵循 timm 标准的 PyTorch 库，通过统一的 create_model() 工厂接口提供多种 3D 点云架构及其预训练权重（如 PointNeXt、Point Transformer、KPConv 等），支持分类、分割、检测和自监督预训练任务。项目仍处于 alpha 阶段，所有模型均经统一重新基准测试，权重托管于 Hugging Face，可 pip 直接安装。
LLM2R casts 3D part assembly as sequence modeling with a pretrained decoder-only LLM, using geometric tokens and learnable summary tokens to predict all part poses in a single non-autoregressive pass, achieving state-of-the-art accuracy on PartNet, PartNeXt, and Breaking Bad benchmarks.
UNITAS是首个3D原生的世界动作模型（WAM），在共享的度量3D坐标系中统一观测、动作与场景动态，以1.7B参数在RoboTwin场景预测上取得最佳结果（位移误差较PointWorld降低最多49%），并在LIBERO上达到99.8%的成功率。
A project report shows that for multi-scan radar object classification on RadarScenes, most of the macro F1 gain (0.7370 → 0.8613) comes from accumulating radar points across scans rather than from the choice of sequence model, with GRU, Transformer, and SSM variants landing in a narrow 0.86–0.89 band, suggesting per-scan representation quality matters more than sequence architecture.
FAMOS is a feed-forward model that predicts movable-part segmentation and joint parameters from sparse point clouds using a Multi-state Articulation Transformer and a procedural data generator, showing consistent improvements over baselines in experiments.
Attention-DP3 enhances 3D diffusion policies by incorporating object-level geometric cues through attention to improve stability in cluttered environments, achieving state-of-the-art performance across benchmarks.
An engineer demonstrates that for a radar-only 5-class object classifier on the RadarScenes dataset, increasing point density from 1 to 5 points per instance roughly doubles macro F1, whereas architectural and feature engineering changes fell within the noise floor. The work highlights how extremely sparse radar points fail to convey size or velocity-spread signatures, causing confusion between classes like two-wheelers and pedestrians.
This paper introduces a fully Bayesian framework for point cloud curve reconstruction using Markov chain Monte Carlo algorithms, which handles noise and missing data while providing uncertainty quantification, with experiments on synthetic and real-world LiDAR data showing accurate results.
A technique for handling sound from irregularly shaped audio sources in game development using point clouds to avoid direction switching issues.
LIDARLearn is an open-source PyTorch library for 3D point cloud deep learning that unifies 56 pre-configured models with built-in cross-validation and automatic publication-ready LaTeX report generation. The framework supports supervised, self-supervised, and parameter-efficient fine-tuning methods across datasets like ModelNet40, ShapeNet, and remote sensing benchmarks.