Diff9D: Diffusion-Based Domain-Generalized Category-Level 9-DoF Object Pose Estimation

TL;DR

This paper tackles domain-generalized category-level $9$-DoF object pose estimation by reframing pose estimation as a diffusion process trained solely on rendered data. It introduces Diff9D, a DDPM-based pipeline with a transformer-based denoiser and a lightweight conditioning module, enabling reverse diffusion in as few as $3$ steps via DDIM for near real-time performance. The method achieves state-of-the-art domain generalization on REAL275 and Wild6D without 3D priors, and demonstrates practical applicability by deploying on a real robot for grasping tasks, achieving about $17.2$ FPS and $80.8\%$ grasp success across categories. Overall, Diff9D advances synthetic-to-real generalization in category-level pose estimation and offers a viable, prior-free solution for real-world robotic manipulation.

Abstract

Nine-degrees-of-freedom (9-DoF) object pose and size estimation is crucial for enabling augmented reality and robotic manipulation. Category-level methods have received extensive research attention due to their potential for generalization to intra-class unknown objects. However, these methods require manual collection and labeling of large-scale real-world training data. To address this problem, we introduce a diffusion-based paradigm for domain-generalized category-level 9-DoF object pose estimation. Our motivation is to leverage the latent generalization ability of the diffusion model to address the domain generalization challenge in object pose estimation. This entails training the model exclusively on rendered synthetic data to achieve generalization to real-world scenes. We propose an effective diffusion model to redefine 9-DoF object pose estimation from a generative perspective. Our model does not require any 3D shape priors during training or inference. By employing the Denoising Diffusion Implicit Model, we demonstrate that the reverse diffusion process can be executed in as few as 3 steps, achieving near real-time performance. Finally, we design a robotic grasping system comprising both hardware and software components. Through comprehensive experiments on two benchmark datasets and the real-world robotic system, we show that our method achieves state-of-the-art domain generalization performance. Our code will be made public at https://github.com/CNJianLiu/Diff9D.

Figures (13)

• Figure 1: Comparison of diffusion model-based image generation and object pose estimation. (a): The process of diffusion model, where ${q}$ and ${p_\theta}$ represent the forward (noising) and reverse (denoising) diffusion processes, respectively. (b): Diffusion model-based image generation task, which generates an image based on a prompt. (c): The Overall pipeline of our Diff9D, which redefines the 9-DoF object pose and size estimation task from a generative perspective, i.e., from Gaussian noise pose to true object pose. ${t_p}$, $R$, and $s$ represent 3D translation, 3D rotation, and 3D size, respectively.
• Figure 2: Some visualizations of the reverse diffusion process, representing the diffusion from Gaussian noise poses to objects poses in the observed scene.
• Figure 3: Workflow of the proposed Diff9D, which includes three main parts (pose diffusion, condition extraction for pose diffusion, and transformer-based denoiser for pose denoising). The input of Diff9D is RGB image, point cloud, and time step $T$ and its corresponding noise pose $Pos{e_T}$. Note that the image is first instance segmented by Mask R-CNN 57 before condition extraction. The condition extraction extracts the input condition $c$. The pose diffusion consists of forward (noising) and reverse (denoising) diffusion processes. Forward diffusion continuously adds noise to the ground-truth object pose $Pos{e_0}$. Reverse diffusion first concatenates the noise pose features ${F_{pose}}$ and $c$ to form the input ${D_{i}}$ for the denoiser. The transformer-based denoiser then takes ${D_{i}}$ as input and predicts the pose noise ${\varepsilon _\theta }\left( {{x_T},T,c} \right)$. Finally, ${\varepsilon _\theta }\left( {{x_T},T,c} \right)$ can be used to denoise $Pos{e_T}$ through the reverse diffusion process based on the Markov chain to obtain $Pos{e_{T - 1}}$. We directly use the translation, size, and rotation matrices to represent the object pose, as shown in Fig. \ref{['Fig1']}. Detailed architecture of the shape estimator and shape encoder is shown in Fig. \ref{['Fig4']}.
• Figure 4: Detailed architecture of the shape estimator and shape encoder. We use two parallel branches to estimate and encode shape and NOCS shape.
• Figure 5: Some visualizations of the estimated shape and NOCS shape. Top, middle, and bottom rows denote the observed RGB images and their corresponding estimated shape and NOCS shape, respectively. It can be seen that these two processes can introduce potential 3D geometric information, making the pose diffusion process geometrically guided.