SurfEmb: Dense and Continuous Correspondence Distributions for Object Pose Estimation with Learnt Surface Embeddings

Project site
Pre-print

Compared to the paper:

• All models are trained for 500k iterations
• 10k pose samples instead of 20k, and maximum 1k pose evaluations after pruning.

We use the available detections from CosyPose, but were not able to find detection-only timings. We add our timing on top of the full CosyPose single-view RGB pipeline, resulting in conservative timings.