AutoRF: Learning 3D Object Radiance Fields from Single View Observations

AutoRF: Learning 3D Object Radiance Fields from Single View Observations

  • 方法
    • Given a single input image, our goal is to encode each 3D object that is present in the scene into a compact representation that allows to, e.g., efficiently store the objects into a database and re-synthesize them from different views/contexts in a later stage.
    • 不需要物体几何先验知识:CAD 模型,对称等
    • 使用预训练的实例或者全景分割找到同一物体的2d像素点,使用预训练的单目3d检测器找到物体3d空间的pose,训练和测试时,每张图可以获取一系列3d框和对应的2dmask,还有相机标定信息
    • 预先准备
      • 图像:先进行单目3d和分割
      • 归一化物体坐标空间:将3dbox进行归一化
      • 以物体为中心的相机
      • occypancy mask: 前景1,背景-1,遮挡或者无法确定0
    • 架构overview
      • 输入:图像,归一化物体坐标空间中的相机(derived by 挖掘物体3d框和mask的信息)
      • 形状和外观编码器:CNN,提出的feature送到平行的两个heads,分别用来生成shape和appearance的code
      • 形状解码器:shape code 被送到解码器网络,可以隐式输出occupancy network,给定一个NOCS里面的3d点可以输出density
      • 外观解码器:同时输入shape和appearance codes,给定3d点和viewing direction输出rgb颜色
      • volume rendering 体渲染 【26】: NERF
    • 训练
      • photometric loss:The difference between the predicted color of the pixel and the actual color of the pixel makes the photometric loss.
      • occupancy loss
    • Test time optimization