Cross attention-based prior deformation for category-level 6D pose estimation

In category-level 6D pose estimation based on shape priors, deforming a canonical shape prior to align with a specific instance is a crucial step. Most existing approaches perform a simple pooling operation to aggregate features from the object instance and use the resulting global features to guide the shape prior's deformation. However, this overly simplified aggregation leads to the loss of important local details, weakening deformation quality and causing inaccuracies in the estimated pose. To address these limitations, we propose a Cross Attention-based Prior Deformation (CAPD) method. CAPD leverages the cross-attention mechanism to explicitly model point-to-point correlations between the shape prior and the instance point cloud. This mechanism enables the propagation of global context from the object instance to the shape prior while also conveying local information, thereby facilitating more precise and expressive deformation. By mitigating the loss of local information, our approach allows the deformed shape prior to better fit the instance geometry, significantly improving robustness in pose estimation. We evaluate CAPD on the real-world REAL275 dataset. Experimental results demonstrate that our method outperforms existing shape-prior-based category-level 6D pose estimation approaches. The code is available at: https://github.com/yangyc275/CAPD_code.