Learning Shape-Independent Transformation via Spherical Representations for Category-Level Object Pose Estimation

TL;DR

This work tackles category-level object pose estimation by addressing semantic incoherence in shape-dependent NOCS coordinates through spherical representations. It introduces SpherePose, which uses a shared proxy sphere discretized by HEALPix to create shape-independent spherical anchors, enabling robust 3D-3D correspondences. The approach centers on three pillars: SO(3)-invariant point-wise features, spherical feature interaction via a Transformer, and a hyperbolic correspondence loss, with translation and size regressed separately. Experiments on CAMERA25, REAL275, and HouseCat6D show state-of-the-art performance and robustness to intra-class shape variation, highlighting the practical impact of spherical representations for category-level pose estimation.

Abstract

Category-level object pose estimation aims to determine the pose and size of novel objects in specific categories. Existing correspondence-based approaches typically adopt point-based representations to establish the correspondences between primitive observed points and normalized object coordinates. However, due to the inherent shape-dependence of canonical coordinates, these methods suffer from semantic incoherence across diverse object shapes. To resolve this issue, we innovatively leverage the sphere as a shared proxy shape of objects to learn shape-independent transformation via spherical representations. Based on this insight, we introduce a novel architecture called SpherePose, which yields precise correspondence prediction through three core designs. Firstly, We endow the point-wise feature extraction with SO(3)-invariance, which facilitates robust mapping between camera coordinate space and object coordinate space regardless of rotation transformation. Secondly, the spherical attention mechanism is designed to propagate and integrate features among spherical anchors from a comprehensive perspective, thus mitigating the interference of noise and incomplete point cloud. Lastly, a hyperbolic correspondence loss function is designed to distinguish subtle distinctions, which can promote the precision of correspondence prediction. Experimental results on CAMERA25, REAL275 and HouseCat6D benchmarks demonstrate the superior performance of our method, verifying the effectiveness of spherical representations and architectural innovations.

Figures (10)

• Figure 1: We propose a novel approach for category-level object pose estimation based on spherical representations. (a) The classical pipeline of correspondence-based methods, where representations denote the organization form of observation data, e.g., point-based representations in $\mathbb{R}^3$, which is straightforward and commonly adopted in previous methods. (b) The overview of comparison between our method and previous methods. Note that each XYZ position of representations and NOCS coordinates is visualized as an RGB tuple. Our method employs spherical representations to learn shape-independent transformation, yielding smaller NOCS angle errors compared to DPDN eccv2022dpdn, which adopts point-based representations and suffers from the shape-dependence.
• Figure 2: Overview of the proposed SpherePose. Given the observation ${\bm{I}}$ and ${\bm{P}}$, we first extract $\mathrm{SO(3)}$-invariant point-wise features from four distinct perspectives and assign them to the spherical anchors ${\bm{A}}$ with HEALPix spherical projection, yielding initial spherical features ${\bm{F}}^A$. Then, the Transformer encoder module is employed for spherical feature interaction and integrates comprehensive spherical features ${\bm{F}}$. Finally, we predict the corresponding spherical NOCS coordinates ${\bm{O}}$ via a NOCS predictor, which is applied to the estimation of rotation ${\bm{R}}$.
• Figure 3: Illustration of several correspondence loss functions and their gradients.
• Figure 4: Overview of (a) equirectangular grids and (b) HEALPix grids.
• Figure 5: PCA visualization of DINOv2 features.