物体姿勢を維持する画像変換を用いたzero-shot姿勢推定

抄録

In this study, we developed a new method to translate unknown products into known products while maintaining the pose of the products, in order to perform pose estimation of unknown products with zero-shot. Pose estimation methods using Neural Networks (NNs) are highly accurate, but their success rate of pose estimation is significantly decreased for unknown products. In retail stores, products are frequently changed, and it is time consuming to prepare data for each new product. Therefore, the purpose in this study is to improve the success rate of pose estimation with zero-shot. Since the success rate of pose estimation are high for known products, unknown products are translated into known products. However, even if only the appearance of the product is translated into the known product, the success rate of pose estimation cannot be improved because the pose of translated appearance often changes. Therefore, the coarse pose of the product is used as input for image translation NN to translate the product into a known product that maintains its pose. Experimental results showed that using pose-maintaining image translation improved the success rate for the two methods of pose estimation which are state-of-the-art on different datasets.