paper_3d__utonia

A cross-domain 3D point cloud encoder produces universal representations across diverse domains like indoor scans and outdoor LiDAR. It uses a PTv3 backbone with domain-aware sampling to enable transfer to segmentation, robotics, and multimodal reasoning.

To transfer 3D point cloud features for robotics manipulation, you use a unified encoder to extract transferable features from your point cloud data. The encoder leverages sparse convolutions and serialization to produce a shared feature space for robotic tasks.

Yes, the PTv3 encoder works with both indoor scans and outdoor LiDAR by applying domain-aware data normalization and grid sampling. This ensures cross-domain alignment across five targeted domains, including CAD models, producing a shared feature space.

To run inference on indoor point cloud data, you install the Utonia encoder and apply the standard transform to your input data. This uses the five-stage PTv3 backbone with sparse convolutions to output transferable 3D representations.

Sparse convolutions are used for 3D point cloud pretraining to enable efficient representation learning across diverse domains. Combined with serialization, they process unstructured point cloud data effectively to produce transferable features for segmentation and reasoning.