Object Pose Distribution Estimation for Determining Revolution and Reflection Uncertainty in Point Clouds

TL;DR

The paper tackles the problem of pose uncertainty in object pose estimation when RGB data are unavailable, focusing on revolution and reflection ambiguities. It introduces a SpyroPose-inspired, 3D-point-cloud method with three components—feature encoder, keypoint feature aggregator, and MLP scorer—to estimate a histogram of pose likelihoods over sampled rotations. The approach achieves 100% precision on predicted ambiguities in both synthetic and real bin-picking scenarios, with practical run-time suitable for near real-time robotics and successful autonomous grasping without extra hardware checks. Extensions to full SE(3) distributions are discussed, along with potential improvements in encoders and multi-modal inputs for broader applicability.

Abstract

Object pose estimation is crucial to robotic perception and typically provides a single-pose estimate. However, a single estimate cannot capture pose uncertainty deriving from visual ambiguity, which can lead to unreliable behavior. Existing pose distribution methods rely heavily on color information, often unavailable in industrial settings. We propose a novel neural network-based method for estimating object pose uncertainty using only 3D colorless data. To the best of our knowledge, this is the first approach that leverages deep learning for pose distribution estimation without relying on RGB input. We validate our method in a real-world bin picking scenario with objects of varying geometric ambiguity. Our current implementation focuses on symmetries in reflection and revolution, but the framework is extendable to full SE(3) pose distribution estimation. Source code available at opde3d.github.io

Figures (8)

• Figure 1: The pose distribution estimate of our method is shown for both unambiguous and ambiguous views. The top images display the test object in two different poses. Below, the distribution estimates are visualized. The red line indicates the probability of each revolution. The dotted green line shows the same revolution probability for the reflected pose. In the left image, the object's indent is visible, allowing for a single object pose to be identified. In the right image, the indent is not visible, resulting in a much larger pose distribution in both revolution and reflection.
• Figure 2:
• Figure 3: Front and back views of the two objects from our experiments. For Object 1 (left) the correct revolution can only be precisely determined from the indent at the bottom, but the reflection can easily be seen from the screw-head. However, for Object 2 (right) both reflection and revolution can only determined from the square recess. Thus in case of self occlusion the correct reflection is ambiguous.
• Figure 4: Synthetic image of Object 2.
• Figure 5: Example of Object 2 in the bins.