HiPose: Hierarchical Binary Surface Encoding and Correspondence Pruning for RGB-D 6DoF Object Pose Estimation (Lin et al.)
June 15, 2026
tldr: fast, dense correspondence method for coarse 6DOF pose estimation from single image
keypoint methods - pick a small, fixed set of distinct points on the object (e.g. 8 corners of a 3D bounding box)
-neural network locates where those points project onto the 2D image
-2D-3D correspondences (2D pixel location —\> known 3D coordinate on the object)
-fed into a PnP solver to determine exact pose (position, orientation)
dense correspondence methods - for every pixel belonging to the object, predict the corresponding 3D coordinate on the object’s surface
-tends to outperform keypoint-based methods because:
1.thousands of correspondences, individual errors get averaged out
2.occlusion handling - if an object is occluded and one of the keypoints happens to fall in that region, you lose a lot of important information
3.keypoints don’t always sit on the actual object, so the network regresses the location of something it doesn’t directly know
iterative closest point - algorithm that finds the rigid transformation (rotation + translation) to align two 3D point clouds
-refinement step, run to polish coarse pose estimates from HiPose, FFB6D, etc.
hierarchical binary decoding
-assigns a bit sequence to each vertex on an object’s mesh
-bits are assigned based on position, and each bit splits a half
-neural network learns to predict bits based on regions instead of memorizing arbitrary mappings to vertices
-used for pose estimation
-uses FFB6D backbone
-instead of predicting point-wise 3D offsets to a keypoint, “classify” the region it belongs to through the learned binary codes