Method:

PoseDETR is an end-to-end, single-stage RGB-only method for direct 6D pose estimation. The approach jointly predicts object instances and their 6D poses. The method does not rely on any default detection or segmentation modules and applies no subsequent iterative pose refinement. A single model per dataset is trained for all objects.

Training Data:

Only provided PBR splits

Used 3D models:

Default object models

Authors:

Temporary Anonymity