DeepHMap++: Combined Projection Grouping and Correspondence Learning for Full DoF Pose Estimation

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Estimating the 6D pose of objects using Convolutional Neural Networks (CNNs) is a significant area of research in computer vision.
• Existing methods are broadly categorized into direct methods and two-stage pipelines, with two-stage pipelines often relying on intermediate cues like 3D coordinates or keypoints.

Purpose of the Study:

• To propose an improved postprocessing framework for two-stage object pose estimation pipelines.
• To enhance object pose estimation in challenging scenes by integrating projection grouping and correspondence learning.

Main Methods:

• A local-patch based method predicts projection heatmaps for 3D bounding box corners.
• A projection grouping module refines these heatmaps by removing redundant local maxima.
• A correspondence evaluation network ranks multiple correspondence hypotheses sampled from heatmaps and their neighborhoods, improving upon direct input to the Perspective-N-Point (PnP) algorithm.

Main Results:

• The proposed framework successfully refines intermediate cues from a two-stage pipeline.
• Experimental results on three public datasets show superior performance compared to several state-of-the-art methods.
• The integration of projection grouping and correspondence learning effectively addresses challenges in pose estimation.

Conclusions:

• The developed framework offers a robust solution for 6D object pose estimation, particularly in complex environments.
• The novel combination of projection grouping and correspondence learning advances the capabilities of two-stage pose estimation pipelines.