Accurate 6D Object Pose Estimation by Pose Conditioned Mesh Reconstruction

Abstract

Current 6D object pose methods consist of deep CNN models fully optimized for a single object but with its architecture standardized among objects with different shapes. In contrast to previous works, we explicitly exploit each object's distinct topological information i.e. 3D dense meshes in the pose estimation model, with an automated process and prior to any post-processing refinement stage. In order to achieve this, we propose a learning framework in which a Graph Convolutional Neural Network reconstructs a pose conditioned 3D mesh of the object. A robust estimation of the allocentric orientation is recovered by computing, in a differentiable manner, the Procrustes' alignment between the canonical and reconstructed dense 3D meshes. 6D egocentric pose is then lifted using additional mask and 2D centroid projection estimations. Our method is capable of self validating its pose estimation by measuring the quality of the reconstructed mesh, which is invaluable in real life applications. In our experiments on the LINEMOD, OCCLUSION and YCB-Video benchmarks, the proposed method outperforms state-of-the-arts.

A review on object pose recovery: From 3D bounding box detectors to full 6D pose estimators

“…While these methods fuse data coming from RGB and depth channels, a local belief propagation based approach [73] and an iterative refinement architecture [31], [32] are proposed in depth modality [74]. 6D object pose estimation is recently achieved from RGB only [164], [170], [172], [173], [174], [30], [37], [38], [40], and the current paradigm is to adopt CNNs [157], [158], [169]. BB8 [40] and Tekin et al [38] perform corner-point regression followed by PnP.…”

Iterative Pose Refinement for Object Pose Estimation Based on RGBD Data

“…Thanks to the rapid development of powerful graphical processing units (GPU), the data-driven techniques have made a great leap in pose estimation [ 8 , 9 ]. Recent methods [ 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 , 18 , 19 , 20 , 21 , 22 ] can be categorized based on the types of input data, i.e., RGB or RGBD. Traditional data-driven approaches [ 23 , 24 ] utilize convolution neural network (CNN) to select candidate feature points that roughly construct a bounding box surrounding the target object in 2D image, and subsequently solve the perspective-n-point (pnp) problem based on these points for pose estimation [ 25 ].…”