Normalized Object Coordinate Space

A shared, standardized coordinate frame for all objects in a category, used to estimate the pose and size of objects never seen before.

NOCS was proposed by He Wang and colleagues at Stanford in a CVPR 2019 paper, for category-level pose estimation — estimating the pose of a specific object instance the model has never seen, within a category it was trained on. The idea is to align and rescale every object in a category into a shared canonical space normalized to a unit cube. Built on top of Mask R-CNN, the network predicts, for every pixel, that pixel’s coordinate within this canonical space (the NOCS map); this is then aligned with the depth map using a similarity transform to recover the object’s 6D pose and 3D size in one step. It requires no CAD model of the specific object, which makes it a representative approach for category-level pose estimation.

ExampleA mug never seen during training is placed on a table; the model segments it and predicts its NOCS map, then combines that with the depth map to compute the mug’s position, orientation, and size.

Also called

NOCS, NOCS Map

Related

Category-Level Pose Estimation · 6D Object Pose Estimation · Instance Segmentation · Mask R-CNN · Depth Camera

Sources

Normalized Object Coordinate Space for Category-Level 6D Object Pose and Size Estimation (arXiv)