HypoPose: Template-Guided Geometric Hypothesis Modeling for Training-Free Unseen 6D Pose Estimation
Abstract
Training-free unseen 6D object pose estimation leverages vision foundation model features to establish correspondences between query and template images and recover object pose without task-specific training or fine-tuning. However, existing methods mainly construct correspondences directly from feature similarity, while overlooking the retrieved template pose itself as an intrinsic spatial constraint; neglecting this geometric reasoning cue limits correspondence reliability and thus degrades pose estimation accuracy. To address this limitation, we propose HypoPose, which explicitly treats the retrieved template pose as a geometric hypothesis to guide correspondence construction. Under each pose hypothesis, rotation-aware similarity is combined with relaxed optimal transport to establish globally consistent correspondences. Subsequent geometric verification prunes inconsistent matches, and multi-hypothesis correspondence fusion aggregates verified matches before final pose estimation. Experiments on the RGB unseen BOP benchmark demonstrate that HypoPose significantly outperforms prior training-free baselines, achieving an Average Recall of 46.2% with a 9.0% absolute improvement. Real-world experiments further illustrate its practical applicability. Code is available at https://github.com/jialeren/hypopose.