arXiv：cs.AI· Zhenyang Liu, Chenjie Cao, Yisu Zhang, Xuhui Zuo, Xiangyang Xue, Yanwei Fu, Tengfei Wang, Chunchao Guo·· 5 小时前
OmniCam：基于几何锚定姿态 token 学习的全景相机轨迹生成
OmniCam: Omni-Camera Trajectory Generation via Geometry-Grounded Pose Token Learning
AI 导读
OmniCam 是一个自回归模型，能从单张全景图和文本轨迹描述生成相机姿态序列，其几何锚定姿态 token 学习整合了全景点云编码器、绝对旋转与相对平移混合 tokenization 以及几何与语义分离的条件流。
正文
Abstract:Camera trajectories control viewpoint changes in video generation, scene reconstruction, and robotic perception. Generating them from language requires both scene geometry and target-aware framing. We introduce OmniCam, an autoregressive model that generates camera pose sequences from a single panorama and textual trajectory descriptions. Its geometry-grounded pose token learning combines three components: a panoramic point-cloud encoder for omnidirectional geometric context; hybrid absolute-rotation and relative-translation tokenization with temporally consistent quaternion signs; and separate geometric and semantic conditioning streams with an explicit 3D target anchor. We also construct OmniCaT, containing 267,700 trajectories across four camera behaviors. On the reported OmniCaT evaluation, OmniCam reduces trajectory errors by 28--47% and collision rate by 65.8% relative to GenDoP retrained on OmniCaT. Against the best baseline for each metric, the ATE and collision reductions are 43.0% and 62.3%, respectively. Component ablations support the use of geometric and target-aware conditioning, while downstream experiments examine camera-controlled video generation and robotic active perception.
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2610.09513 [cs.CV] |
| (or arXiv:2610.09513v2 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2610.09513 arXiv-issued DOI via DataCite |
Submission history
From: Zhenyang Liu [view email]
[v1]
Wed, 7 Oct 2026 06:08:28 UTC (32,437 KB)
[v2]
Sat, 10 Oct 2026 03:35:33 UTC (32,437 KB)
来源：arXiv：cs.AI · arxiv.org