I2L-MeshNet: Image-to-Lixel Prediction Network for Accurate 3D Human Pose and Mesh Estimation from a Single RGB Image

Most of the previous image-based 3D human pose and mesh estimation methods\nestimate parameters of the human mesh model from an input image. However,\ndirectly regressing the parameters from the input image is a highly non-linear\nmapping because it breaks the spatial relationship between pixels in the input\nimage. In addition, it cannot model the prediction uncertainty, which can make\ntraining harder. To resolve the above issues, we propose I2L-MeshNet, an\nimage-to-lixel (line+pixel) prediction network. The proposed I2L-MeshNet\npredicts the per-lixel likelihood on 1D heatmaps for each mesh vertex\ncoordinate instead of directly regressing the parameters. Our lixel-based 1D\nheatmap preserves the spatial relationship in the input image and models the\nprediction uncertainty. We demonstrate the benefit of the image-to-lixel\nprediction and show that the proposed I2L-MeshNet outperforms previous methods.\nThe code is publicly available https://github.com/mks0601/I2L-MeshNet_RELEASE.\n