Multi-View Object Pose Refinement With Differentiable Renderer

This paper introduces a novel multi-view 6 DoF object pose refinement approach focusing on improving methods trained synthetic data. It is based the DPOD detector, which produces dense 2D-3D correspondences between model vertices and image pixels in each frame. We have opted for use of multiple frames with known relative camera transformations, as it allows introduction geometrical constraints via an interpretable ICP-like loss function. The function implemented differentiable renderer optimized iteratively. also demonstrate that full detection pipeline, solely data, can be used auto-labeling real perform quantitative evaluation LineMOD, Occlusion, Homebrewed YCB-V datasets report excellent performance comparison to state-of-the-art empirically our requires only few robust close locations noise extrinsic calibration, making its practical usage easier more ubiquitous.