Hyperparameter-Free Losses for Model-Based Monocular Reconstruction

Abstract

This work proposes novel hyperparameter-free losses for single view 3D reconstruction with morphable models (3DMM). We dispense with the hyperparameters used in other works by exploiting geometry, so that the shape of the object and the camera pose are jointly optimized in a sole term expression. This simplification reduces the optimization time and its complexity. Moreover, we propose a novel implicit regularization technique based on random virtual projections that does not require additional 2D or 3D annotations. Our experiments suggest that minimizing a shape reprojection error together with the proposed implicit regularization is especially suitable for applications that require precise alignment between geometry and image spaces, such as augmented reality. We evaluate our losses on a large scale dataset with 3D ground truth and publish our implementations to facilitate reproducibility and public benchmarking in this field.

Multi-View 3D Face Reconstruction in the Wild Using Siamese Networks

“…This is often translated on appending multiple terms into the loss together with hyperparameters that help to keep the norm ofα id small [19]. In order to avoid the use of hyperparameters, we make use of the loss proposed in [13], which allows to simultaneously learn the 3D shape and the camera pose using a sole term expression.…”

“…In order to enforce a global scene consistency in the predictions, we define an objective that uses all the camera poses and the predicted 3D shape within a single term, which does not include any hyperparameter and is easy to minimize, as proposed by [13]. Our loss is defined as the sum of the reprojection errors across all the input views, as in Bundle Adjustment [21], which is the Maximum Likelihood estimator when the image error is zero-mean.…”

“…Finally, it is not trivial how to combine information from multiple views in order to satisfy the geometric constrains of the scene and generate better reconstructions as the number of views increases. Recent approaches introduce geometric inductive biases about the scene into the architectures [10] and the losses [13], which constrain the number of possible solutions and ease the learning process.…”