Temporal Consistent Object Pose Estimation from Monocular Videos

The existing object pose estimation methods mostly rely on single-frame observations, which exhibit poor robustness in occluded scenes. Since obtaining multi-frame sequential inputs is plausible in robot operations, we present a framework to perform object pose estimation from a monocular sequence. Different from predicting poses in each view before fusing them, we propose to directly reconstruct a volumetric representation for each sequence to provide temporal consistent pose estimates. A gated recurrent units module is deployed to guide the network to fuse features from previous and current views. The experiments conducted on widely recognized benchmarks demonstrate that our approach surpasses the performance of state-of-the-art techniques.