StanfordExtra

从单目互联网图像中恢复狗的 3D 姿势和形状的端到端方法。犬种之间的巨大形状差异、显着的遮挡和低质量的互联网图像使这成为一个具有挑战性的问题。我们比以前的工作学习了更丰富的先验形状，这有助于规范参数估计。我们在斯坦福狗数据集上展示了结果，这是一个包含 20,580 张狗图像的“野外”数据集，我们已经收集了 2D 关节和轮廓注释以进行分割以进行训练和评估。为了捕捉狗的各种形状，我们表明 2D 数据集中的自然变化足以通过期望最大化 (EM) 学习详细的 3D 先验。作为培训的副产品，我们生成了一个新的参数化模型（包括肢体缩放）SMBLD，我们将其与我们的新注释数据集 StanfordExtra 一起发布给研究社区。An end-to-end approach for recovering 3D pose and shape of dogs from monocular Internet images. Vast shape differences among dog breeds, significant occlusions, and low-quality Internet images make this a highly challenging problem. We learn richer shape priors than previous works, which helps regularize parameter estimation. We demonstrate our results on the Stanford Dogs Dataset, a "in-the-wild" dataset containing 20,580 dog images, for which we have collected 2D joint and contour annotations for segmentation, training and evaluation. To capture the diverse shapes of dogs, we show that the natural variations present in the 2D dataset are sufficient to learn detailed 3D priors via Expectation-Maximization (EM). As a byproduct of training, we generate a new parametric model (including limb scaling) named SMBLD, which we are releasing to the research community alongside our newly annotated dataset, StanfordExtra.

搜集汇总

背景与挑战

背景概述

StanfordExtra是一个用于狗3D姿势和形状恢复的计算机视觉数据集，基于斯坦福狗数据集的20,580张图像，并提供了2D关节和轮廓注释。该数据集旨在通过端到端方法学习先验形状，以应对狗种差异和图像质量挑战，并发布了新的参数化模型SMBLD供研究使用，由微软·剑桥大学于2021年发布。