Search Results

Now showing 1 - 3 of 3
  • Item
    Neural Denoising for Spectral Monte Carlo Rendering
    (The Eurographics Association, 2022) Rouphael, Robin; Noizet, Mathieu; Prévost, Stéphanie; Deleau, Hervé; Steffenel, Luiz-Angelo; Lucas, Laurent; Sauvage, Basile; Hasic-Telalovic, Jasminka
    Spectral Monte Carlo (MC) rendering is still to be largely adopted partially due to the specific noise, called color noise, induced by wavelength-dependent phenomenons. Motivated by the recent advances in Monte Carlo noise reduction using Deep Learning, we propose to apply the same approach to color noise. Our implementation and training managed to reconstruct a noise-free output while conserving high-frequency details despite a loss of contrast. To address this issue, we designed a three-step pipeline using the contribution of a secondary denoiser to obtain high-quality results.
  • Item
    Fast and Fine Disparity Reconstruction for Wide-baseline Camera Arrays with Deep Neural Networks
    (The Eurographics Association, 2022) Barrios, Théo; Gerhards, Julien; Prévost, Stéphanie; Loscos, Celine; Sauvage, Basile; Hasic-Telalovic, Jasminka
    Recently, disparity-based 3D reconstruction for stereo camera pairs and light field cameras have been greatly improved with the uprising of deep learning-based methods. However, only few of these approaches address wide-baseline camera arrays which require specific solutions. In this paper, we introduce a deep-learning based pipeline for multi-view disparity inference from images of a wide-baseline camera array. The network builds a low-resolution disparity map and retains the original resolution with an additional up scaling step. Our solution successfully answers to wide-baseline array configurations and infers disparity for full HD images at interactive times, while reducing quantification error compared to the state of the art.
  • Item
    3D Human Shape and Pose from a Single Depth Image with Deep Dense Correspondence Enabled Model Fitting
    (The Eurographics Association, 2022) Wang, Xiaofang; Boukhayma, Adnane; Prévost, Stéphanie; Desjardin, Eric; Loscos, Celine; Multon, Franck; Sauvage, Basile; Hasic-Telalovic, Jasminka
    We propose a two-stage hybrid method, with no initialization, for 3D human shape and pose estimation from a single depth image, combining the benefits of deep learning and optimization. First, a convolutional neural network predicts pixel-wise dense semantic correspondences to a template geometry, in the form of body part segmentation labels and normalized canonical geometry vertex coordinates. Using these two outputs, pixel-to-vertex correspondences are computed in a six-dimensional embedding of the template geometry through nearest neighbor. Second, a parametric shape model (SMPL) is fitted to the depth data by minimizing vertex distances to the input. Extensive evaluation on both real and synthetic human shape in motion datasets shows that our method yields quantitatively and qualitatively satisfactory results and state-of-the-art reconstruction errors.