MAPRPose: Mask-Aware Proposal and Amodal Refinement for Multi-Object 6D Pose Estimation

2026-04-22 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Computer Vision & Pattern Recognition · Depth: Expert, quick

Summary

MAPRPose is a novel two-stage framework designed for multi-object 6D pose estimation, specifically addressing challenges posed by severe occlusion and sensor noise in cluttered scenes. The first stage, Mask-Aware Pose Proposal (MAPP), establishes reliable 3D keypoint matches by lifting 2D correspondences into 3D space, generating geometrically consistent pose hypotheses. It then selects the top-K candidates based on correspondence-level scoring. The subsequent refinement stage incorporates an Amodal Mask Prediction and ROI Re-Alignment (AMPR) module within a tensorized render-and-compare pipeline. AMPR reconstructs complete object geometry and dynamically adjusts the Region-of-Interest to reduce localization errors and spatial misalignment under heavy occlusion. MAPRPose also features GPU-accelerated RGB-XYZ reprojection, allowing simultaneous refinement of multiple pose hypotheses. On the BOP benchmark, MAPRPose achieved an Average Recall (AR) of 76.5%, surpassing FoundationPose by 3.1% AR and demonstrating a 43x speedup in multi-object inference.

Key takeaway

For research scientists developing 6D object pose estimation systems, MAPRPose offers a significant advancement in handling occlusion and improving inference speed. You should consider integrating mask-aware proposal generation and amodal refinement techniques into your pipelines to achieve higher accuracy and efficiency, especially in complex, cluttered environments. This approach demonstrates superior performance and substantial speedup compared to prior methods.

Key insights

MAPRPose enhances 6D pose estimation by combining mask-aware proposals with amodal refinement for robust occlusion handling.

Principles

• Leverage mask-aware correspondences for robust pose proposals.
• Reconstruct complete object geometry for amodal refinement.

Method

MAPRPose uses a two-stage process: Mask-Aware Pose Proposal (MAPP) for initial hypotheses, followed by Amodal Mask Prediction and ROI Re-Alignment (AMPR) for robust refinement via render-and-compare.

In practice

• Utilize 3D keypoint matches for geometrically consistent pose.
• Employ dynamic ROI adjustment to mitigate occlusion errors.

Topics

• 6D Pose Estimation
• Multi-Object Pose
• MAPRPose
• Mask-Aware Proposal
• Amodal Refinement

Best for: Research Scientist, AI Scientist, Computer Vision Engineer, Machine Learning Engineer

Related on AIssential

• Amodal SAM: A Unified Amodal Segmentation Framework with Generalization — Amodal SAM extends SAM's generalization to predict complete object shapes, including occluded regions, using synthetic data and novel learning.
• Exploring 6D Object Pose Estimation with Deformation — DeSOPE dataset enables 6D pose estimation for deformed objects, revealing existing methods struggle with non-rigid shapes.
• ReCoSplat: Autoregressive Feed-Forward Gaussian Splatting Using Render-and-Compare — ReCoSplat uses a Render-and-Compare module and KV cache compression for robust, online novel view synthesis.
• Adaptive Slicing-Assisted Hyper Inference for Enhanced Small Object Detection in High-Resolution Imagery — ASAHI adaptively slices high-resolution images to enhance small object detection, reducing computation and improving performance.
• Plan2Map: A Multimodal Benchmark for Document-Grounded Geospatial Boundary Reconstruction from Planning Records — Plan2Map enables multimodal geospatial boundary reconstruction from diverse planning documents via structured processing.
• PanoPlane: Plane-Aware Panoramic Completion for Sparse-View Indoor 3D Gaussian Splatting — Panoramic scene completion with layout-anchored attention steering significantly improves sparse-view 3D reconstruction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.