AuraTracer智迹闻
中文

EVENT DOSSIER

Learning Spherical Occupancy Profiles for Multi-View 3D Reconstruction and Generation

2026-09-07 12:00 Science 🔥 40.2 heat score
1sources
1days unfolding
40.2heat score
2mentions
SummaryAI generated

On September 7, 2026, arXiv cs.CV published the paper “Learning Spherical Occupancy Profiles for Multi-View 3D Reconstruction and Generation”. This study proposes a new method based on spherical occupancy profiles, aimed at improving the ability of 3D reconstruction and content generation in multi-view scenarios.

Related eventsRELATED EVENTS
Key entitiesKEY ENTITIES
Google Scanned ObjectsarXiv

Coverage · reports per dayLANGUAGE SPLIT

Entity relations
Google Scanned Objects …1

SignalsSIGNALS

Keyword heat
  • arXiv1
  • Google Scanned Objects1

All reports (1)SOURCES

A arXiv cs.CV en 2026-09-07 12:00

Learning Spherical Occupancy Profiles for Multi-View 3D Reconstruction and Generation

研究人员提出球形占用剖面作为统一中间表示,用于从图像进行判别式和生成式 3D 重建。在包含 999 个物体的 Google Scanned Objects 数据集子集(每个物体 48 张旋转台视图)上训练了判别式逐射线解码器和基于剖面 VAE 及潜在扩散模型的生成流水线。判别式模型将全局视图平均和射线特定图像证据注入 FiLM 条件化的剖面头,在独立 90 个物体的测试集上达到中值软深度误差 0.035(归一化)。生成流水线支持无条件采样以匹配重建流形,以及通过分类器自由引导可量化且可调的基于图像的多元解重建。后验功率锐化和学习到的锐化目标均能恢复地面真剖面宽度而不降低深度,揭示了 L1-逐射线损失族中的单调宽度 - 峰值前沿并推动了形态门的原则性重新定义。在两个 DTU 场景上的真实照片验证证实该流程可迁移…