Skip to content

Latest commit

 

History

History
27 lines (23 loc) · 2.33 KB

codes_explanations.md

File metadata and controls

27 lines (23 loc) · 2.33 KB

We have provided some explanation on our codes. Welcome for any questions or discussions.

Explanations on variable names

  • que means "query", which is the "test view" in the paper. e.g. que_imgs_info contains information about the test views.
    • que_imgs_info['Ks'] has the size qn*3*3 is the intrinsics matrices of the test views. qn means the number of test views in this batch. (qn means "query number")
    • que_imgs_info['poses'] has the size qn*3*4 is the pose matrices $[R;t]$ of the test views. We use the opencv-style poses which converts the scene coordinate to the camera coordinate $x_{cam}= Rx_{scene}+t$.
    • que_imgs_info['depth_range'] has the size qn*2, which are the near plane depth, and the far plane depth.
    • que_imgs_info['coords'] has the size qn*rn*2, which are rn 2D coordinates in pixel on test views. We will render the rays emitted from these coordinates. (rn means "ray number")
    • que_depth has the size qn*rn*dn and is the sample depth values on test rays. dn means the number of points sampled on a test ray. ($K_t$ in the paper)
  • ref means "reference", which is the "input view" in the paper. e.g. ref_imgs_info contains information about the input views.
    • ref_imgs_info['ray_feats'] has the size rfn*f*h*w, which is the visibility feature map $G$ on input views. rfn means the "reference view number", i.e. the number of input views (working view number $N_w$ in the paper). f means the dimension number. h*w is the size of this feature map.
  • prj means information about projected sample points on input views. e.g. prj_dict
  • nr means "network rendering", which is computed from the constructed radiance fields. e.g. pixel_colors_nr means the output colors computed by volume rendering on the constructed radiance field.
  • dr means "direction rendering", which is directly computed from the NeuRay representation.
  • Summary of matrix size in annotations
    • qn test view number
    • rn test ray number
    • rfn input working view number $N_w$
    • dn sample point number on a test ray $K_t$
    • f feature dimension
    • pn=qn*rn*dn total sample point number

Dataset management

All datasets are managed by BaseDatabase in dataset/database.py. If we want to extend to a new dataset, we can write a new subclass of BaseDatabase and implement all its functions.