Flow6D: Discrete-to-Continuous Flow Matching for Efficient and Accurate Category-Level 6D Pose Estimation

Existing 6D pose estimation methods often struggle to balance accuracy and computational efficiency. Direct regression in continuous space is prone to local optima and noise, while candidate-based search methods are computationally expensive. Flow6D addresses this by proposing a hierarchical, two-stage approach that transitions from discrete latent localization to continuous pose refinement.

Flow6D decomposes the pose estimation task into two distinct stages:

1. Discrete Latent Localization: The rotation and translation parameters are discretized into bins. A discrete flow matching (DFM) model learns to predict the probability distribution over these bins, effectively narrowing the search space to high-confidence regions and avoiding the pitfalls of direct continuous regression.
2. Continuous Pose Regression: Using the initial estimate from the first stage, a continuous flow matching (CFM) model predicts fine-grained pose residuals. By sampling within a localized residual space, the model optimizes the pose to achieve high precision, effectively bridging the gap between discrete binning and exact coordinates.

This framework is designed to be efficient, utilizing deterministic transport paths to minimize inference time, and is adaptable to both rigid and articulated objects.

Flow6D demonstrates superior performance on both synthetic and real-world datasets (such as REAL275). It consistently outperforms existing state-of-the-art methods in accuracy metrics while maintaining a high inference speed of 70 FPS. The hierarchical design successfully mitigates the search complexity of infinite continuous spaces, providing a robust foundation for precise pose recovery even in complex, occluded, or partially observed scenes.