R_relative and T_relative are bascilly rigid transform on the camera over the exisiting base transformation. Bascially it transform the original coordinate axis of the camera to new coordinate axis and the rotation and translation between the two are given by R_relative and T_relative
Rendering Point Clouds from RGB-D Images (10 points)
Parametric Functions (10 + 5 points)
New Shape Elipsoid
Implicit Surfaces (15 + 5 points)
New Shape Elipsoid (Voxel size = 256). Trade off vs a Parametric representation : (Speed) For meshes its proportional to the faces while for point clouds its proportional to the number of points. Generally number of faces are more in number than points. However in practive rendering meshes is faster due to modern algorithm like z-biffering. (Memory) Meshes have to store both face connectivity and vertices, while point clouds have to only store points. Which technically means meshes have higher memory footprint. However to match the rendering quality we need higher number of points so they are comparable in natuere. (rendering quality) The rendering quality of meshes are obiously better due to C0 level continuity. Point clouds are sparse and one can never fully infer the surface topology from just by observing point cloud. (ease of use) Points cloud seems to be better since infering about connecivity is non trivial and hence many paper still stuggle to predict meshes directly but there are many works that are able to directly predict point clouds.
