Resolving Symmetry Ambiguity in Correspondence-based Methods for Instance-level Object Pose Estimation

Yongliang Lin, Yongzhi Su, Sandeep Inuganti, Yan Di, Naeem Ajilforoushan, Hanqing Yang, Yu Zhang, Jason R. Rambach

Introduction

Estimating the 6D pose (3D location and 3D orientation) of an object from a single RGB image is a fundamental challenge in robotics and augmented reality. Most high-performing methods rely on finding correspondences, attempting to match 2D image pixels to specific 3D coordinates on an object's surface. However, "Symmetry Ambiguity" occurs when objects look identical from multiple angles—such as a plain bowl, a bolt, or a sphere. In these cases, a single pixel could correspond to multiple valid points on the 3D surface. Traditional "one-to-one" matching algorithms often fail here, leading to training instability and requiring computationally expensive geometric solvers like PnP-RANSAC to filter out erroneous matches.

SymCode: Symmetry-aware Surface Encoding

To address this, researchers have developed SymCode, a novel mapping system that explicitly accounts for symmetry. Instead of forcing a network to select a single unique 3D point for every pixel, SymCode employs a binary grouping scheme. It identifies all points on a 3D model that are "symmetrically equivalent" and assigns them the same code.

Comparison of different surface encodings: (c) traditional coordinates, (d) ZebraPose, and (e) the proposed SymCode, which handles discrete and continuous symmetries more effectively.

This approach transforms the task from a difficult regression problem into a stable classification problem. SymCode is designed to handle various types of symmetry, including discrete (e.g., a square peg), continuous (e.g., a cylinder), and even objects with minor irregularities that are "almost" symmetrical. By embracing the inherent ambiguity of symmetric objects rather than trying to resolve it prematurely, the network's learning objective becomes significantly more consistent.

SymNet and the CPR Module

Beyond the encoding, the authors introduce SymNet, a direct regression network that eliminates the need for traditional geometric solvers. Most correspondence methods use a Perspective-n-Point (PnP) algorithm to calculate the final pose from matches, but PnP is often slow and sensitive to outliers.

The SymNet architecture, featuring a Binary Encoding Module to predict SymCodes and a Correspondence-based Pose Regression (CPR) module for direct 6D pose estimation.

SymNet utilizes a Correspondence-based Pose Regression (CPR) module. This neural network takes predicted SymCode maps and segmentation masks as input and directly regresses the 6D pose parameters. This bypasses the PnP step entirely, resulting in a system that is both faster and more robust against occlusions.

Experimental Results

The method was evaluated on two challenging benchmarks: T-LESS, which features 30 textureless, industry-relevant objects with various symmetries, and IC-BIN, which focuses on bin-picking scenarios with heavy occlusion.

  • Speed: SymNet achieves an inference time of approximately 0.093 seconds per image, making it roughly 3x faster than previous state-of-the-art methods like ZebraPose (0.25s) that rely on PnP.
  • Accuracy: On the T-LESS dataset, SymNet demonstrated a 25% improvement in object recall compared to earlier symmetry-handling baselines.
  • Robustness: Qualitative analysis shows the model handles cluttered environments and severe occlusions more effectively than one-to-one methods because it is not forced to identify a unique point that may be hidden from view.

Conclusion and Future Work

SymCode and SymNet represent a significant step forward in instance-level object pose estimation by treating symmetry as a feature rather than a bug. While the current pipeline requires a manual step to partition complex models and depends on an initial 2D bounding box detector, it sets a new standard for speed and accuracy in symmetric object handling. Future work may focus on automating the partitioning process and extending this symmetry-aware logic to category-level estimation where exact CAD models are unknown.

Summary updated

BibTeX
@article{lin2025resolving,
  author = {Yongliang Lin and Yongzhi Su and Sandeep Inuganti and Yan Di and Naeem Ajilforoushan and Hanqing Yang and Yu Zhang and Jason R. Rambach},
  title = {{Resolving Symmetry Ambiguity in Correspondence-based Methods for Instance-level Object Pose Estimation}},
  journal = {IEEE Trans. Image Process.},
  year = {2025},
  doi = {10.1109/tip.2025.3544142},
}
Figure