Search Results

Now showing 1 - 2 of 2
  • Item
    Simple Techniques for a Novel Human Body Pose Optimisation Using Differentiable Inverse Rendering
    (The Eurographics Association, 2022) Battogtokh, Munkhtulga; Borgo, Rita; Pelechano, Nuria; Vanderhaeghe, David
    Human body 3D reconstruction has a wide range of applications including 3D-printing, art, games, and even technical sport analysis. This is a challenging problem due to 2D ambiguity, diversity of human poses, and costs in obtaining multiple views. We propose a novel optimisation scheme that bypasses the prior bias bottleneck and the 2D-pose annotation bottleneck that we identify in single-view reconstruction, and move towards low-resource photo-realistic 3D reconstruction that directly and fully utilises the target image. Our scheme combines domain-specific method SMPLify-X and domain-agnostic inverse rendering method redner, with two simple yet powerful techniques. We demonstrate that our techniques can 1) improve the accuracy of the reconstructed body both qualitatively and quantitatively for challenging inputs, and 2) control optimisation to a selected part only. Our ideas promise extension to more difficult problems and domains even beyond human body reconstruction.
  • Item
    Fitness of General-Purpose Monocular Depth Estimation Architectures for Transparent Structures
    (The Eurographics Association, 2022) Wirth, Tristan; Jamili, Aria; Buelow, Max von; Knauthe, Volker; Guthe, Stefan; Pelechano, Nuria; Vanderhaeghe, David
    Due to material properties, monocular depth estimation of transparent structures is inherently challenging. Recent advances leverage additional knowledge that is not available in all contexts, i.e., known shape or depth information from a sensor. General-purpose machine learning models, that do not utilize such additional knowledge, have not yet been explicitly evaluated regarding their performance on transparent structures. In this work, we show that these models show poor performance on the depth estimation of transparent structures. However, fine-tuning on suitable data sets, such as ClearGrasp, increases their estimation performance on the task at hand. Our evaluations show that high performance on general-purpose benchmarks translates well into performance on transparent objects after fine-tuning. Furthermore, our analysis suggests, that state-of-theart high-performing models are not able to capture a high grade of detail from both the image foreground and background at the same time. This finding shows the demand for a combination of existing models to further enhance depth estimation quality.