Towards Human-Level 3D Relative Pose Estimation: Generalizable, Training-Free, With Single Reference.

Area of Science:

• Computer Vision
• Robotics
• 3D Perception

Background:

• Humans intuitively estimate object pose from single images.
• Existing methods often require extensive training data and object-specific labeling.
• Leveraging 3D shape perception, render-and-compare, and semantic cues is key.

Purpose of the Study:

• To propose a novel 3D generalizable relative pose estimation method.
• To enable pose estimation for unseen objects without prior training or labeling.
• To improve upon existing supervised methods using a training-free approach.

Main Methods:

• Utilizes a 2.5D shape from an RGB-D reference image.
• Employs a differentiable renderer for a render-and-compare simulation.
• Leverages semantic cues from pre-trained models (e.g., DINOv2) for correspondence.
• Refines 3D relative pose by comparing rendered and query images/semantic maps.

Main Results:

• The proposed method achieves state-of-the-art performance on LineMOD, LM-O, and YCB-V datasets.
• Demonstrates significant outperformance over supervised methods, particularly with rigorous accuracy metrics (Acc@5/10/15°).
• Shows strong generalization capabilities in challenging cross-dataset experiments.

Conclusions:

• The training-free, generalizable relative pose estimation method is effective for unseen objects.
• The integration of 2.5D shape, differentiable rendering, and semantic cues is a promising direction.
• This approach offers a robust alternative to supervised methods in 3D pose estimation tasks.